THE LOCAL STRUCTURE OF

SMOOTH MAPS OF MANIFOLDS

By

Jonathan Michael Bloom

SUBMITTED IN PARTIAL FULFILLMENT OF THE

REQUIREMENTS FOR THE DEGREE OF

BACHELORS OF ARTS WITH HONORS

AT

HARVARD UNIVERSITY

CAMBRIDGE, MASSACHUSETTS

APRIL 2004

Table of Contents

Table of Contents i

Acknowledgements iii

Introduction 1

1 Preliminaries: Smooth Manifolds 3

1.1 Smooth Manifolds and Smooth Maps . . . . . . . . . . . . . . . . . . 3

1.2 Linearization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2 Power: The Theorem of Sard 14

2.1 Critical Values and Regular Values . . . . . . . . . . . . . . . . . . . 14

2.2 Proof of Sard’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . 17

3 Language: Transversality 21

3.1 The Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.2 Jet Bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

3.3 The Whitney C∞ Topology . . . . . . . . . . . . . . . . . . . . . . . . 34

3.4 The Thom Transversality Theorem . . . . . . . . . . . . . . . . . . . 37

3.5 The Multijet Transversality Theorem . . . . . . . . . . . . . . . . . . 42

4 High to Low: Morse Theory 44

4.1 The Morse Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

4.2 Morse Functions are Generic . . . . . . . . . . . . . . . . . . . . . . . 50

5 Low to High: The Whitney Embedding Theorem 54

5.1 Embeddings and Proper Maps . . . . . . . . . . . . . . . . . . . . . . 54

5.2 Proof of the Whitney Embedding Theorem . . . . . . . . . . . . . . . 58

i

ii

6 Two to Two: Maps from the Plane to the Plane 62

6.1 Folds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

6.2 Cusps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

Bibliography 71

Acknowledgements

I am profoundly grateful to my advisor, Professor Peter Kronheimer. Without his

skillful guidance, this thesis would not have been possible. Although my topic is not

his primary field of study, he always knew the next important question to be asked

and connection to be explored; time and again he saved me from losing my way. On

a personal level, I am thankful for his compassionate supervision and the confidence

he placed in me. I would also like to thank Professor Barry Mazur, who graciously

met with me long ago to discuss possible thesis topics. I walked straight from that

meeting to the library and buried myself in a stack of books on Singularity Theory. I

could not have hoped for a more satisfying topic of study. I am indebted to Professor

Wilfried Schmid for possessing the boldness to define objects the right way from the

start and the pedagogical talent to do so effectively. After four years of thought, I

can say that I understand most of what he had hoped to convey. I would never have

had this opportunity in the first place had it not been for the inspirational efforts of

the late Professor Arnold Ross, who encouraged me to think deeply of simply things.

I am deeply indebted to Peter Anderegg, Benjamin Bakker, and Gopal Sarma,

each of whom selflessly devoted many hours to proofreading my drafts. I would also

like to thank my parents and grandparents for their loving support and grand care

packages; my roommates Chris, Christine, James, Jenny, Monica, Nico, Ryan, Wendy,

and especially Daniela, who was a caring friend during the spring break; Ameera,

George, and Jennifer for always checking in on me; Owen and Nick for seeing the big

picture; Sasha and Dimitar for their thesis camaraderie; Dr. John Boller for spurring

a torrent of writing; George and David for LATEXsupport; Geraldine for providing me

with endless electronica and trip-hop; and Olga for keeping life exciting.

Cambridge, MA Jonathan Bloom

April 5, 2004

iii

Introduction

Functions, just like human beings,

are characterized by their singularities.

- P. Montel.

The aim of this thesis is to provide a self-contained introduction to the modern study

of the local structure of smooth maps of manifolds. By smooth, we mean differentiable

to all orders. We consider two smooth maps to have the same local structure at a

point if the maps are locally equivalent up to a change of coordinates. A critical

point of a smooth map is a point of the domain at which the derivative does not

have full rank. In §1 we show that, away from critical points, a smooth map exhibits

a single, simple local structure (that of a linear projection). Thus the remaining

chapters are concerned with the structure of smooth maps at their critical points,

i.e. with the structure of singularities. In general, infinitely many local structures

can occur at singularities. However, with certain restrictions on dimension, almost

all maps exhibit at most a select handful of these. Our goal, then, is to discern which

types of singularities can generically occur, and then to establish normal forms for

the structure of these singularities.

1

2

The first three chapters are devoted to developing machinery to this end, with an

emphasis on transversality. In §1, we provide the necessary background on smooth

manifolds and prove the aforementioned result on the local structure of maps away

from singularities. In §2, we prove Sard’s Theorem, a fundamental piece of analysis

that will power the more elaborate machinery of §3. In §3, we introduce the notion

of transversality and prove a variety of elementary results. We then construct the jet

bundle of smooth maps of manifolds and use it to place a topology on the space of

smooth maps. Finally, we prove and then generalize our fundamental tool: the Thom

Transversality Theorem. This theorem is a potent formalization of the intuition that

almost all maps are transverse to a fixed submanifold.

The latter three chapters then use these tools to accomplish our goal in the three

simplest situations, respectively. In §4, we look at smooth functions (by which we

mean smooth maps of manifolds down to R). This is the situation studied in Morse

Theory. We will see that the generic singularities occur at so-called non-degenerate

critical points and then establish normal quadratic forms for smooth functions at such

points. In §5, we consider the opposite extreme, where the codomain has at least

twice the dimension of the domain. We will see that in these dimensions there are no

generic singularities, and then we will extend this result to prove the famous Whitney

Embedding Theorem: Every smooth n-manifold can be embedded into R2n+1. Lastly,

in §6, we look at smooth maps between 2-manifolds and show that generically the

singularities consist of folds along curves with isolated cusps. We then establish

normal forms for folds and simple cusps.

Chapter 1

Preliminaries: Smooth Manifolds

In §1.1, we present the necessary background on smooth manifolds and smooth maps.

In §1.2, we give normal forms for the local structure of smooth maps at regular points.

1.1 Smooth Manifolds and Smooth Maps

The reader who is already familiar with smooth manifolds is advised to skim this

section. Our treatment is adapted from Lang and the notation is standard [7]. We

define manifolds as abstract objects having certain properties rather than as special

subsets of Euclidean space. The latter approach, taken in [6] and [14], has the advan-

tage of concreteness. We often think of manifolds this way, and it is perhaps best as

a first exposure. In §5, we will see that the abstract definition of a manifold is in fact

no more general: any manifold can be realized as a submanifold of some Euclidean

space. Since the concrete definition is equivalent, one might wonder why we bother

with the abstract definition in the first place.

In fact, the abstract approach echoes a general theme in higher mathematics: we

define (or redefine) a category of object by its relevant properties and then prove that

3

4

any object with these properties is in fact in the category. By proceeding abstractly,

we can define the object with as little extraneous data as possible. The concrete

approach to manifolds fails in this regard since it equips each manifold with a non-

canonical embedding. To see why this excess baggage causes difficulty, consider the

task of constructing new manifolds. When an abstract manifold does not arise natu-

rally as a subset of Euclidean space, it is often far simpler to verify that it satisfies

the properties of the abstract definition than to find a real embedding.

So without further ado, let X be set.

Definition 1.1.1. A smooth atlas on X is a collection of pairs (Ui, ϕi) satisfying the

following conditions for some fixed n:

(1) Each Ui is a subset of X and the Ui cover X.

(2) Each ϕi is a bijection of Ui onto an open subset ϕi(Ui) of Rn and ϕi(Ui

⋂Uj)

is open in Rn for each pair i, j.

(3) The map

ϕj · ϕ−1i : ϕi(Ui

⋂Uj) → ϕj(Ui

⋂Uj)

is a smooth diffeomorphism for each pair i, j.

We can give X a topology in a unique way such that each Ui is open and each ϕi

is a homeomorphism. This topology inherits the properties of local compactness and

second countability from Euclidean space. In the theory of manifolds, we will also

require this topology to be Hausdorff and, thus, paracompact.

Each pair (Ui, ϕi) is called a chart of the atlas. If x ∈ Ui then we call (Ui, ϕi)

a chart at x. Let U be an open subset of X and ϕ a homeomorphism from U onto

some open subset of Rn. Then the pair (U,ϕ) is compatible with the atlas (Ui, ϕi) if

5

ϕ · ϕ−1i is a diffeomorphism for each i such that U

⋂Ui 6= ∅. Two atlases on X are

compatible if every chart of the first is compatible with every chart of the second. It

is straightforward to check that this gives an equivalence relation on the collection of

atlases.

Definition 1.1.2. A smooth manifold is a set X together with an equivalence class

of atlases on X, called its differential structure. A smooth manifold is n-dimensional

if each chart maps into Rn.

A chart ϕ : U → Rn is given by n coordinate functions ϕ1, ..., ϕn, often written

x1, ..., xn. Under ϕ, a point p ∈ U has image (x1(p), ..., xn(p)), or, abusing notation,

simply (x1, ..., xn). We call the functions x1, ..., xn a set of local coordinates on U , and

sometimes we will refer to U as a coordinate neighborhood (or just ‘nbhd’).

Given smooth manifolds X and Y with atlases (Ui, ϕi) and (Vi, ψi), respectively,

we can equip the set X × Y with the atlas (Ui × Vj, ϕi × ψj). This manifold is the

product of X and Y .

Now let X be a smooth manifold and Z a subset of X. Suppose that for each

point z ∈ Z there exists a chart (U,ϕ) at z such that:

(1) ϕ gives a diffeomorphism of U with a product V1 × V2 where V1 is an open

subset of Rm and V2 is an open subset of Rp, with m and p fixed.

(2) ϕ(Z⋂

U) = V1 × v2 for some v2 ∈ V2.

It is not difficult to verify that the collection of pairs (Z⋃

U,ϕ|Z ⋃U) constitutes a

smooth atlas for Z. The set Z equipped with this differential structure is called a

submanifold of X. If follows that at every point of a submanifold Z, we can find a

chart sending Z ⊂ X diffeomorphically to Rk ⊂ Rn. Also note that an open subset of

X can be given the structure of an open submanifold, and a submanifold of a strictly

6

lower dimension is necessarily closed in X. The codimension of a submanifold Z of

X is defined by the equation codim Z = dim X − dim Z.

We use the differential structure on a manifold to define the notion of a smooth

map of manifolds.

Definition 1.1.3. Let X and Y be two smooth manifolds. Let f : X → Y be a map.

(1) f is smooth if, given x ∈ X, there exists a chart (U,ϕ) at x and and a chart

(V, ψ) at f(x) such that f(U) ⊂ V and the map ψ · f ·ϕ−1 : ϕ(U) → ψ(V ) is smooth.

(2) By C∞(X,Y ) we denote the space of smooth maps from X to Y .

(3) f is a diffeomorphism if it is smooth and has a smooth inverse.

In general, there is no canonical choice of local coordinates near a point of a

manifold. The space of maps seems blurred, as distinct maps of manifolds can yield

identical maps of Euclidean spaces when pulled back and pushed forward via distinct

charts. We can make this notion of equivalence of maps precise up to a “change of

coordinates” without invoking charts at all.

Definition 1.1.4. Let f, g : X → Y be smooth maps.

(1) f and g are equivalent if there exist diffeomorphisms h1 of X and h2 of Y such

that the following diagram commutes:

f

X - Y

h1

? ?

h2

X - Y

g

7

(2) f and g are locally equivalent at a point x ∈ X if there exists an open neigh-

borhood U of x such that f |U and g|U are equivalent. Here U is considered as an

open submanifold.

With f and g as above, let (U,ϕ) and (V, ψ) be charts at x and g(x) respectively

with f(U) ⊂ V . Then if f and g are locally equivalent at x, we sometimes say (even

though it is technically not true) that f is locally equivalent at x to the map ψ·g·ϕ−1 of

Euclidean spaces. This second use of the term “local equivalence” proves convenient

for statements about the normal forms of maps and should not cause confusion in

practice. Note that both equivalence and local equivalence are equivalence relations

on the space of smooth maps from X to Y , denoted C∞(X, Y ).

We would like the derivative of a smooth map of manifolds to be a smooth map

of manifolds as well. To this end, we introduce the notions of the tangent space and

tangent bundle of a smooth manifold.

Definition 1.1.5. Let X be an n-dimensional smooth manifold and x ∈ X. Consider

triples of the form (U,ϕ, v), where (U,ϕ) is a chart at x and v ∈ Rn.

(1) Two triples (U,ϕ, v) and (V, ψ, w) are equivalent if

d(ψ · ϕ−1)ϕ(x)v = w.

This gives an equivalence relation on such triples.

(2) A tangent vector to X at x is an equivalence class of such triples.

(3) The tangent space to X at x, denoted TxX, is the set of all tangent vectors to

X at x. We can equip TxX with the topology of the vector space Rn via the bijection

sending the equivalence class of (U,ϕ, v) to v, with (U,ϕ) a fixed chart at x.

8

(4) The tangent bundle of X is the set TX =⊔

x∈X TxX (disjoint union). We can

canonically construct an equivalence class of atlases on TX in order to give TX the

structure of a 2n-dimensional smooth manifold.

Let f : X → Y be a smooth map of manifolds. Using charts we can interpret the

derivative of f at x as the unique map (df)x : TxX → Tf(x)Y having the following

property: If (U,ϕ) is a chart at x and (V, ψ) is a chart at f(x) such that f(U) ⊂ V ,

and if ~v = (U,ϕ, v) is a tangent vector at x represented by v in the chart (U,ϕ), then

(df)x(~v) is the tangent vector at f(x) represented by d(ψ ·f ·ϕ−1)ϕ(x)(v). Fixing these

charts, we have the diagram

(df)x

TxX - Tf(x)Y

? ?Rn - Rm

d(ψ · f · ϕ−1)ϕ(x)

From the diagram it is clear that (df)x is linear. Recall that the corank of a linear

map λ : Rn → Rm is defined to be min{n,m} − rank λ. If corank λ = 0 we say that

λ has full rank.

Definition 1.1.6. Let f : X → Y be a smooth map of manifolds and x ∈ X.

The rank (corank) of f at x is the rank (corank) of (df)x as a linear map from

TxX to Tf(x)Y .

We will need one more elementary result on smooth manifolds [14, p.35].

9

Theorem 1.1.1 (Inverse Function Theorem). Let X and Y be smooth manifolds

of equal dimension, f : X → Y a smooth map, and x ∈ X such that f has full rank

at x. Then there exists an open nbhd U of x such that f(U) is an open nbhd of f(x)

and f |U has a smooth inverse.

The Inverse Function Theorem suggests that the rank of f at x encodes a good

deal of information about the structure of f near x. We will begin to freely exploit

this connection in the next section.

1.2 Linearization

The goal of this section is to completely determine the local structure of a smooth

map at any point of full rank (up to local equivalence). At such a point, a smooth

map must be at least one of the following types.

Definition 1.2.1. Let f : X → Y be a smooth map of manifolds with q = f(p).

(1) f is an immersion at p if (df)p : TpX → TqY is injective.

(2) f is a submersion at p if (df)p : TpX → TqY is surjective.

(3) f is a local diffeomorphism at p if (df)p : TpX → TqY is bijective.

(1′) f is an immersion if f is an immersion at p for every p ∈ X.

(2′) f is an submersion if f is an submersion at p for every p ∈ X.

(3′) f is an local diffeomorphism if f is a local diffeomorphism at p for every p ∈ X.

The following lemma immediately follows from the above definitions and the fact

that dim X = dim TpX.

Lemma 1.2.1. Let f : X → Y be a smooth map of manifolds.

10

(1) If f is an immersion at any point then dim X ≤ dim Y .

(2) If f is a submersion at any point then dim X ≥ dim Y .

(3) If f is a local diffeomorphism at any point then dim X = dim Y .

Lemma 1.2.1 suggests the logic behind the names immersion and submersion. The

former places a smaller manifold into a larger one while the latter smoothly packs a

larger manifold into a smaller one. It is important to note, however, that the image of

either type of map need not be a submanifold. In §5, we will see examples of this and

establish sufficient conditions for the image of a smooth map to be a submanifold.

We now construct normal forms for the local structure of immersions, submersions,

and local diffeomorphisms. In fact, most of the work of this is encoded by the Inverse

Function Theorem. To make this more explicit, we now express this theorem using

the concept of local equivalence (in the sense following its definition in §1.1). The

reader is encouraged to verify that the following theorem is simply a restatement of

Theorem 1.1.1.

Theorem 1.2.2 (Inverse Function Theorem). Let f : X → Y be a smooth map

of manifolds. Suppose that f is a local diffeomorphism at a point p ∈ X. Then f is

locally equivalent to the identity map (x1, ..., xn) 7→ (x1, ..., xn) at p.

The Inverse Function Theorem allows us to locally linearize maps between mani-

folds of different dimension as well. We will first show this in the case of immersions

(so necessarily dim X ≤ dim Y ).

Lemma 1.2.3 (Immersion Lemma). Let f : X → Y be a smooth map of manifolds.

Suppose f is an immersion at a point p ∈ X. Then f is locally equivalent to the map

(x1, ..., xn) 7→ (x1, ...xn, 0, ..., 0) at p.

11

Proof. Our strategy is to construct a map between equidimensional manifolds and

then apply the Inverse Function Theorem. f is a immersion at p so locally we may

assume f : Rn → Rn × Rl, where l = m− n. Then

(df)p =

( ∂f1

∂x1)p ( ∂f1

∂x2)p · · · ( ∂f1

∂xn)p

......

...

(∂fm

∂x1)p (∂fm

∂x2)p · · · (∂fm

∂xn)p

has rank n, where fi is the ith coordinate function of f . Thus there exist n linearly

independent rows of (df)p. Without loss of generality we assume the first n rows are

linearly independent (otherwise we can permute the coordinates on the range). Define

f : Rn → Rn by f = (f1, ...fn), so det(df)p 6= 0.

Now define F : Rn × Rl → Rn × Rl by F (x, y) = f(x) + (0, y). Then

(dF )p =

(df)p 0

∗ Il

so det(dF )p = det(df)p 6= 0. By the Inverse Function Theorem, F has a local smooth

inverse G near p. On a neighborhood of p we have G · f(x) = G · F (x, 0) = (x, 0).

Thus f is linearized by a change of coordinates on the range.

Next, we establish a normal form for submersions (so necessarily dim X ≥ dim Y ).

Lemma 1.2.4 (Submersion Lemma). Let f : X → Y be a smooth map of man-

ifolds. Suppose f is a submersion at a point p ∈ X. Then f is locally equivalent to

the map (x1, ..., xn) 7→ (x1, ..., xm) at p.

Proof. Again our strategy is to construct a map between equidimensional manifolds

and then apply the Inverse Function Theorem. f is an immersion at p so locally we

may assume f : Rm × Rl → Rm, where l = n−m. Then

12

(df)p =

( ∂f1

∂x1)p ( ∂f1

∂x2)p · · · ( ∂f1

∂xn)p

......

...

(∂fm

∂x1)p (∂fm

∂x2)p · · · (∂fm

∂xn)p

has rank m, where fi is the ith coordinate function of f . Thus there exist m

linearly independent columns of (df)p. Without loss of generality we assume the

first m columns are linearly independent (otherwise we can permute the coordinates

on the domain). Define f : Rm → Rm by f(x1, ..., xm) = f(x1, ..., xm, 0, ...0), so

det(df)p 6= 0.

Now define F : Rm × Rl → Rm × Rl by F (x, y) = (f(x, y), y). Then

(dF )p =

(df)p ∗0 Il

so det(dF )p = det(df)p 6= 0. By the Inverse Function Theorem, F is a diffeomorphism

near p. Let g : Rm×Rl → Rm be the projection onto Rm. Then on a neighborhood of

p we have g · F (x, y) = g(f(x, y), y) = f(x, y). Thus f is linearized to g by a change

of coordinates on the domain.

We can combine these last two results to give a complete characterization of the

local structure of smooth maps at points of full rank.

Theorem 1.2.5 (Linearization Theorem). Let f : X → Y be a smooth map of

manifolds. Let p ∈ X be a point such that (df)p has full rank. Then f is locally

equivalent to any linear map of full rank at p.

Such a point p is called a regular point. We have established that the local

structure of smooth functions at regular points is as simple as we could have hoped.

A natural question to ask is: Just how common are regular points? It turns out that

13

this is not the “right” question to ask. In the next chapter, we will both formulate

the “right” question and provide a very precise answer with Sard’s Theorem.

Chapter 2

Power: The Theorem of Sard

In §2.1, we show that smooth maps pull regular points back to submanifolds. In §2.2,we prove Sard’s Theorem.

2.1 Critical Values and Regular Values

Definition 2.1.1. Let f : X → Y be a smooth map of manifolds.

(1) p ∈ X is a critical point of f if corank (dfp) > 0.

(2) p ∈ X is a regular point of f if corank (dfp) = 0.

(3) q ∈ Y is a critical value of f if there exists a critical point p of f with q = f(p).

(4) q ∈ Y is a regular value of f if q is not a critical value of f .

Note that the regular values of f include all points of Y − f(X).

We ended §1.2 with a question: How common are regular points? Our intuition

from calculus suggests that most smooth maps have many regular points and few

critical points. Still, there exists a very simple map for which every point is critical,

namely a constant map. The key observation is that while a constant map has many

critical points, it has only one critical value. Near a critical point, the degeneration of

14

15

the derivative limits the range of the map in a way made precise by Taylor’s Theorem.

Informally, Sard’s theorem states that the set of critical values of a smooth map is

small. We now make precise this notion of “small”.

Definition 2.1.2. (1) Let S be a subset of Rn. Then S has measure zero if for every

ε > 0, there exists a cover of S by a countable number of open cubes C1, C2, ... such

that∑∞

i=1 vol[Ci] < ε.

(2) Let X be a smooth manifold and S a subset of X. Then S is of measure zero

if there exists a countable open cover U1, U2, ... of S and charts φi : Ui → Rn such

that φi(Ui

⋂S) has measure zero in Rn.

Sard’s Theorem says that the set of critical values of a smooth map of manifolds

has measure zero in Y . We will prove Sard’s Theorem in the next section. For now

we establish the most useful property of regular values.

Theorem 2.1.1. Let f : X → Y be a smooth map of manifolds, q ∈ Y a regular

value of f . Then f−1(q) is a smooth submanifold of X with dim f−1(q) = dim X −min{dim X, dim Y }.

Proof. Let q ∈ Y be a regular value of f , W = f−1(q), n = dim X, and m = dim Y .

Suppose p ∈ W .

If n ≤ m then f is an immersion at p. By the Immersion Lemma, we can choose

local coordinates on X and Y so that p = q = 0 and f(x1, ..., xn) = (x1, ..., xn, 0, ..., 0).

Thus p is an isolated point of W ⊂ X (i.e., there exists a nbhd of p containing no

other point of W ). Since the topology of X is second countable, we conclude that W

consists of countable number of isolated points and is therefore a smooth submanifold

of dimension 0.

16

If n ≥ m then f is a submersion at p. By the Submersion Lemma, we can choose

local coordinates on X and Y so that p = q = 0 and f(x1, ..., xn) = (x1, ..., xm). Then

locally W = {(0, ..., 0, xm+1, ..., xn)} so xm+1, ...xn serve as local coordinates for W

near p, mapping a nbhd of p in W (under the subspace topology) diffeomorphically

onto the subspace Rn−m ⊂ Rn.

Remark 2.1.1. If q is not in the image of f then W = ∅, which vacuously satisfies the

definition of a manifold of any dimension.

Theorem 2.1.1 provides a useful tool for generating manifolds. For example, we

can realize the n-sphere as a submanifold of Rn+1 by considering the preimage of 0

under the map (x1, ..., xn+1) 7→ 1− (x21 + · · ·+ x2

n+1).

The same method used to prove Theorem 2.1.1 yields a related result that we will

need in §3.1.

Proposition 2.1.2. Let f : X → Y be a map of smooth manifolds, and W a sub-

manifold of Y . If f is a submersion at each point in f−1(W ), then f−1(W ) is a

submanifold of X with codim f−1(W ) = codim W .

Proof. Let l = dim W . Suppose p ∈ X with f(p) = q. First choose local coordinates

y1, ..., ym on Y so that, near q, W corresponds to the first l coordinates. Next choose

local coordinates x1, ..., xn on X to linearize f near p to f(x1, ..., xn) = (x1, ..., xm).

Then locally f−1(W ) = {(x1, ..., xl, 0, ..., 0, xm+1, ..., xn)} so x1, ..., xl, xm+1, ...xn serve

as local coordinates for f−1(W ) near p, mapping a nbhd of p in f−1(W ) (under the

subspace topology) diffeomorphically onto the subspace Rn−(m−l) ⊂ Rn.

17

2.2 Proof of Sard’s Theorem

Theorem 2.2.1 (Sard). Let f : X → Y be a smooth map of manifolds. Then the

set of critical values of f has measure zero in Y .

It suffices to prove the following:

Proposition 2.2.2. Let U be an open subset of Rn, and let f : U → Rm be a smooth

map. Then the set of critical values of f has measure zero in Rm.

Proof of Theorem from the Proposition. Let C denote the set of critical points of f .

Let {Ui} be a countable open cover of X with the property that each f(Ui) is contained

in some coordinate nbhd Vi of Y with chart ϕi : Vi → Rm. By Proposition 2.2.2, the

set of critical values of ϕi ·f |Uihas measure zero in Rm. So ϕi(Vi

⋂f(C)) has measure

zero in Rm. Therefore f(C) is of measure zero in Y .

In order to carry out the induction step in our proof of the Proposition 2.2.2, we

will need the the following result from measure theory [15, p.51]:

Theorem 2.2.3 (Fubini). Let S be a subset of Rn = R× Rn−1 such that the inter-

section of S with each hyperplane x × Rn−1 has measure zero in Rn−1. Then S has

measure zero in Rn.

We will also need the elementary fact that the countable union of measure zero

sets has measure zero. Our proof follows that of Milnor [11, p.16].

Proof of Proposition 2.2.2. We proceed by induction on n. The case n = 0 holds

trivially. Assume true for n − 1. Let C be the set of critical points of f . For each

i ≥ 1, let Ci be the set of points x ∈ X such that all of the partial derivatives of f of

order ≤ i vanish at x. Clearly the Ci form a descending chain

18

C ⊃ C1 ⊃ C2 ⊃ C3 · · · .

We divide the proof into parts:

Part A: f(C − C1) has measure zero.

Part B: f(Ci − Ci+1) has measure zero for every i ≥ 1.

Part C: f(Ck) has measure zero for some k.

Assuming the above, f(C) = f(C−C1)⋃ (⋃k−1

i=1 f(Ci − Ci+1)) ⋃

f(Ck) is a count-

able union of measure zero sets. Therefore f(C) has measure zero.

Proof of Part A: For each x ∈ (C − C1), we will show that there is an open nbhd

V of x such that f(V⋂

C) has measure zero. A countable number of such nbhds will

cover Rn, so we conclude that f(C − C1) is a countable union of measure zero sets.

So let x ∈ (C − C1). Then the first order partial derivatives of f at x are not all

zero, so we may assume without loss of generality that ∂f1

∂x1(x) 6= 0. Define the map

h : U → Rn by

h(x1, ..., xn) = (f1(x1, ..., xn), x2, ..., xn).

By construction (dh)x is invertible. By the Inverse Function Theorem, there exists

an open nbhd V ∈ U of x such that h|V is a diffeomorphism onto its image W . Now

consider the map g = f · h−1 : W → Rm. Let C ′ ⊂ W be the set of critical points

of g. Since h−1 is a diffeomorphism, h−1(C ′) = V⋂

C and thus g(C ′) = f(V⋂

C).

So it suffices to show that g(C ′) has measure zero. Since h−1 = (f−11 , ...), the first

component function of g = f · h−1 is the identity map. Thus, for fixed x1 ∈ R, g

maps the hyperplane {x1} × Rn−1 into the hyperplane {x1} × Rm−1. By induction,

g(C ′ ⋂{x1}×Rn−1) has measure zero in {x1}×Rm−1 for every x1 ∈ R. Therefore by

Fubini’s Theorem, g(C ′) has measure zero.

19

Proof of Part B: Again it suffices to show that, for each x ∈ (Ci − Ci+1), there is

an open nbhd V of x such that f(V⋂

C) has measure zero.

Let x ∈ (Ci−Ci+1). Then ∂i+1fr

∂xs1∂xs2 ···∂xsi+1(x) 6= 0 for some 1 ≤ s1, ..., si+1 ≤ n and

the map

w(x) = ∂i+1fr

∂xs2 ···∂xsi+1(x)

vanishes on Ci but ∂w∂xs1

6= 0. We may assume s1 = 1. Define h : U → Rn by

h(x1, ..., xn) = (w1(x1, ..., xn), x2, ..., xn).

Note that h carries Ci to the hyperplane 0× Rn−1. As before (dh)x is invertible and

there exists an open nbhd V ∈ U of x such that h|V is a diffeomorphism onto its

image W . Again consider the map g = f · h−1 : W → Rm. By induction, the set of

critical values of the restriction

g : 0× Rn−1⋂

W → Rm

has measure zero in Rm. But since h carries Ci to 0 × Rn−1, we have (Ci

⋂V ) ⊂

h−1(0 × Rn−1). Therefore the set of critical values of g contains f(Ci

⋂V ), so

f(Ci

⋂V ) has measure zero as well.

Proof of Part C: Let In ⊂ U be a cube of edge length δ. We will show that for k

sufficiently large, f(Ck

⋂In) has measure zero in Rm. Then, since we can cover Ck

with a countable number of such cubes, f(Ck) has measure zero as well.

By Taylor’s Theorem, the compactness of In, and the definition of Ck, we see that

f(x + h) = f(x) + R(x, h)

where

‖R(x, h)‖ ≤ c ‖h‖k+1

20

for x ∈ Ck

⋂In and x+h ∈ In. Here c ∈ R depends only on f and In. Now subdivide

In into rn cubes of edge length δr. Suppose I1 is a sub-cube containing a point x ∈ Ck.

It follows from induction on the Pythagorean Theorem that the diameter of I1 is√

n δr.

Therefore any point of I1 can be expressed as x+h with ‖h‖ ≤ √n δ

r. From the Taylor

formula and the triangle inequality, it follows that f(I1) sits inside a cube with edge

length ark+1 , where a = 2(

√nδ)k+1. Thus f(Ck

⋂In) is contained in a union of at

most rn sub-cubes with total volume

V ≤ rn( ark+1 )

m = amrn−(k+1)m.

Hence if k > nm− 1, we can cover f(Ck

⋂In) with a countable number of cubes of

arbitrarily small total volume by choosing r to be sufficiently large (i.e., by taking

a sufficiently fine partition of In into sub-cubes). Therefore f(Ck

⋂In) has measure

zero.

It is a basic result of measure theory that a measure zero subset of Rn cannot

contain a non-empty open set. This fact clearly extends to manifolds, so we have...

Corollary 2.2.4. Let f : X → Y be a smooth map of manifolds. Then the regular

values of f form a dense subset of Y .

Chapter 3

Language: Transversality

In §3.1, we define the basic notions of the theory of transversality and show that trans-

verse maps pull submanifolds back to submanifolds. In §3.2, we define the jet bundle

and use it to topologize C∞(X,Y ) in §3.3. In §3.4, we prove the Thom Transversality

Theorem and generalize to the Multijet Transversality Theorem in §3.5.

3.1 The Basics

Transversality can be viewed as a far-reaching generalization of the notion of regular

value in §2. Informally, two manifolds are transverse if their intersection occurs in

the most general possible form. This is best conveyed through a series of images

(see the next page). Notice that the intersections in each of the the non-transverse

images can be remove or made transverse with only a slight adjustment to one of

the manifolds. By contrast, in each of the images labeled transverse, the intersection

cannot be easily removed. In this sense, transverse intersections are stable.

21

22

Figure 3.1: [6, p.30]

Figure 3.2: [6, p.31]

23

We can neatly define transversality of manifolds using tangent spaces.

Definition 3.1.1. Let X be a smooth manifold, W and Z submanifolds. Then W

and Z are transverse at p ∈ X, denoted W tp Z, if either:

a) p /∈ W⋂

Z

b) p ∈ W⋂

Z and TpX = TpW + TpZ.

W and Z are transverse, denoted W t Z, if W tp Z for every p ∈ X.

The reader is encouraged to review the figures with this definition in mind. In

order to study the local structure of smooth maps, we must also define what it means

for a smooth map to be transverse to a submanifold.

Definition 3.1.2. Let f : X → Y be a smooth map of manifolds, W a submanifold

of Y . Then f is transverse to W at p, denoted f tp W , if either:

(a) f(p) /∈ W

(b) f(p) ∈ W and Tf(p)Y = (df)p(TpX) + Tf(p)W

f is transverse to W, denoted f t W , if f tp W for every p ∈ X. We may also write

f t W on U ⊂ X to convey that f tp W for every p ∈ U .

Remark 3.1.1. (1) Suppose that dim X ≥ dim Y and that W consists of a single point

q ∈ Y . Then f t W iff q is a regular value of f .

(2) It also follows immediately that submersions are transverse to every subman-

ifold.

(3) Some texts define f t W iff graph(f) t X ×W [4, p.39]. The reader should

verify that these two definitions are equivalent.

Imagine a smooth curve in R2 intersecting itself transversely at a point. With

only two dimensions in which to move, it is impossible to remove this intersection

24

through an arbitrarily small perturbation. However, if we now embed the curve in

R3, we can remove the intersection with ease:

Figure 3.3: [6, p.50]

In R3 the intersection is not stable, so it cannot be transverse. Transverse curves

in R3 must not intersect at all, and the next proposition generalizes this intuition to

higher dimensional manifolds.

Proposition 3.1.1. Let f : X → Y be a map of smooth manifolds, W a submanifold

of Y . Suppose that dim X + dim W < dim Y . Then f t W iff the image of f is

disjoint from W .

Proof. The latter condition clearly implies the former.

Now suppose there exists p ∈ X with f(p) ∈ W . Then

dim[(df)p(TpX) + Tf(p)W ] ≤ dim(df)p(TpX) + dim Tf(p)W

≤ dim X + dim W

< dim Y = dim Tf(p)Y

Therefore f is not transverse to W at p.

In order to prove the main result of this section, we will need the following lemma.

25

Lemma 3.1.2. Let f : X → Y be a map of smooth manifolds, W a submanifold

of Y of codimension k, and p ∈ X with f(p) ∈ W . Let U be a nbhd of f(p) and

φ : U → Rk a submersion such that W⋂

U = φ−1(0). Then f t W at p iff φ · f is a

submersion at p.

Proof. Since φ is a submersion at f(p), dim im((dφ)f(p)) = k = codim W . Thus

dim ker(dφ)f(p) = dim W = dim Tf(p)W . And Tf(p)W ⊂ ker(dφ)f(p) because φ is

constant on W . Therefore ker(dφ)f(p) = Tf(p)W .

Now φ ·f is a submersion at p iff (dφ ·f)p is onto. Since (dφ)f(p) is onto, the latter

condition holds

iff (df)p(TpX) + ker((dφ)f(p)) = Tf(p)Y

iff (df)p(TpX) + Tf(p)W = Tf(p)Y

iff f t W at p.

By Remark 3.1.1, we can view the next theorem as a generalization of Theorem

2.1.1.

Theorem 3.1.3. Let f : X → Y be a map of smooth manifolds, W a submanifold of

Y . If f t W then f−1(W ) is a submanifold of X with codim f−1(W ) = codim W .

Proof. By the definition of submanifold, it is sufficient to show that every p ∈ f−1(W )

has a neighborhood V ⊂ X such that f−1(W )⋂

V is a submanifold of X. Suppose p ∈X with f(p) = q. Since W is a submanifold, there is a neighborhood U ⊂ Y of q and

a submersion φ : U → Rk (with k = dim W ) such that W⋂

U = φ−1(0). By Lemma

3.1.2, f t W at p implies that φ·f is a submersion at p. Therefore, by the Submersion

Lemma, we can choose a sufficiently small neighborhood V ⊂ X of p such that

f(V ) ⊂ U and φ·f a submersion on V . Near p we have f−1(W )⋂

V = (φ·(f |V ))−1(0)

which is a submanifold of the stated dimension by Proposition 2.1.2.

26

As an immediate corollary, we see that the intersection of transverse manifolds is

itself a manifold.

Corollary 3.1.4. Let X be a smooth manifold, W and Z submanifolds. If W t Z

then W⋂

Z is a submanifold of X with codim (W⋂

Z) = codim W + codim Z.

Proof. Let i : Z → X be the inclusion map. Then (di)p(TpZ) = TpZ, so W t Z

implies i t Z. W⋂

Z = i−1(Z) so we are done by Theorem 3.1.3.

We can make any two manifolds transverse through an arbitrarily small pertur-

bation of one of them. The following lemma formalizes this intuition and is the key

to proving the more powerful transversality theorems of §3.4 and §3.5. Note how its

proof translates the power of Sard’s Theorem into a statement about transversality.

Lemma 3.1.5 (Transversality Lemma). Let X, B, and Y be smooth manifolds, W

a submanifold of Y . Let j : B → C∞(X,Y ) be a set map and define Φ : X × B → Y

by Φ(x, b) = j(b)(x). If Φ t W then the set {b ∈ B | j(b) t W} is dense in B.

Proof. Let WΦ = Φ−1(W ). Since Φ t W , WΦ is a submanifold of X × B. Let

π : WΦ → B be the restriction of the projection X × B → B. If b /∈ π(WΦ), then

j(b)(x) /∈ W for every x ∈ X, so j(b)(X)⋂

W = ∅ and j(b) t W .

If dim WΦ < dim B, then π(WΦ) has measure zero in B by Corollary 2.2.4. There-

fore j(b) t W for every b in the dense set B − π(WΦ).

For the case dim WΦ ≥ dim B, it suffices to prove the following claim and apply

Corollary 2.2.4.

Claim: If b is a regular value of π, then j(b) t W .

Proof of Claim: Let b be a regular value of π and x ∈ X. We will show that j(b) tx

W . If (x, b) /∈ WΦ, then j(b)(x) /∈ W , so j(b) tx W . So assume (x, b) ∈ WΦ. Since b

27

is a regular value of π and dim WΦ ≥ dim B, we have that (dπ)(x,b)(T(x,b)WΦ) = TbB,

or that T(x,b)WΦ contains T(x,b)({x} ×B) as subsets of T(x,b)X ×B. Thus

T(x,b)(X ×B) = T(x,b)WΦ + T(x,b)(X × {b})

Applying (dΦ)(x,b) to both sides gives

(dΦ)(x,b)(T(x,b)(X ×B)) = Tj(b)(x))W + (dj(b))x(TxX)

By assumption Φ t W so

TΦ(x,b)Y = TΦ(x,b)W + (dΦ)(x,b)(T(x,b)X ×B)

Combining gives

Tj(b)(x)Y = Tj(b)(x)W + (dj(b))x(TxX)

Therefore j(b) t W at x.

As our first application of the Transversality Lemma, we show that we can make

a map transverse to a submanifold with an arbitrarily small translation.

Proposition 3.1.6. Let f : X → Rm be a smooth map of manifolds, W a submanifold

of Rm. Then {b ∈ Rm | (f + b) t W} is a dense subset of Rm.

Proof. In the notation of the Transversality Lemma, let B = Y = Rm and define

j : Rn → C∞(X,Rn) by j(b) = f + b. Then Φ : X × B → B defined by Φ(x, b) =

j(b)(x) = f(x) + b is clearly a submersion. Therefore Φ t W and we are done by

Proposition 2.1.2.

Once we have placed a topology on C∞(X,Y ), we will be able to make further use

of the Transversality Lemma to prove more powerful claims about the set of maps

transverse to a submanifold. We develop this topology in the next two sections.

28

3.2 Jet Bundles

We can define the jet bundles of smooth maps of manifolds with coordinates or ab-

stractly. In [13], Saunders describes the first jet bundle in terms of local coordinates

in order to best study first order differential equations, which can be interpreted

as closed embedded submanifolds of this bundle [13, p.103]. However, as one pro-

gresses to higher order jet bundles, the coordinate-based approach becomes decidedly

unwieldy. For our purposes, it makes sense to follow the invariant treatment of jet

bundles in [5]. Advantages include notational simplicity, increased utility, and greater

overall beauty. Though the definition may appear strange at first, we will soon de-

scend to local coordinates in order to develop an intuition for this new structure.

Definition 3.2.1. Let f, g : X → Y be smooth maps of manifolds. Let p ∈ X with

f(p) = g(p) = q.

(1) f has first order contact with g at p if (df)p = (dg)p as maps of TpX → TqY .

(2) f has kth order contact with g at p if (df) : TX → TY has (k − 1)th order

contact with (dg) at every point in TpX. We denote this by f ∼k g.

(3) Let Jk(X,Y )p,q denote the set of equivalence classes under“∼k at p” of maps

f : X → Y with f(p) = q.

(4) Let Jk(X, Y ) =⊔

(p,q)∈X×Y Jk(X, Y )p,q. Jk(X,Y ) is the kth jet bundle of maps

from X → Y and an element σ of Jk(X, Y ) is called a k-jet of maps from X → Y .

(5) Let σ ∈ Jk(X, Y ). Then there exist unique p ∈ X and q ∈ Y with σ ∈Jk(X, Y )p,q. p is called the source of σ and q the target of σ. The map α : Jk(X, Y ) →X defined by σ 7→ (source of σ) is called the source map. The map β : Jk(X, Y ) → Y

defined by σ 7→ (target of σ) is called the target map.

29

To each smooth map f : X → Y we can assign a map jkf : X → Jk(X, Y )

sending a point x to the k-jet with source p by f , also called the k-jet of f at p. Note

that J0(X, Y ) ∼= X × Y and j0f(X) is simply the graph of f .

The k-jet of f at p is just an invariant way of describing the Taylor expansion of

f at p to kth order:

Proposition 3.2.1. Let U be an open neighborhood of p in Rn, f, g : U → Rm smooth

mappings, fi and gi the coordinate functions of f and g, respectively, and x1, ..., xn

coordinates on U . Then f ∼k g at p iff

∂|α|fi

∂xα (p) = ∂|α|gi

∂xα (p)

for every multi-index α with |α| ≤ k and 1 ≤ i ≤ m.

Proof. We proceed by induction on k. For k = 1, we have f ∼1 g at p iff (df)p = (dg)p

iff ∂fi

∂xj(p) = ∂gi

∂xj(p) for every 1 ≤ i, j ≤ n.

Assume the proposition is true for k − 1. Let y1, y2, ... be the coordinates of Rn

in U × Rn = TU . Then (df) : U × Rn → Rm × Rm = TRm is given by

(x, y) 7→ (f(x), f 1(x, y), ..., fm(x, y))

where

f i(x, y) =∑n

j=1∂fi

∂xj(x)yj

And similarly for (dg).

If f ∼k g at p then (df)p ∼k−1 (dg)p on TpX. By the induction hypothesis, (df)

and (dg) have the equal partial derivatives to (k− 1)st order at each (p, v) ∈ p×Rn.

So for each multi-index α with |α| ≤ k − 1, we have

∂|α|f i

∂xα (p, v) = ∂|α|gi

∂xα (p, v)

30

Evaluating at v = (0, ..., 1, ...0) (1 is the jth position) gives

∂|α|∂xα

∂fi

∂xj(p) = ∂|α|

∂xα∂gi

xj(p)

All partial derivatives of f and g order ≤ k are obtained this way.

Conversely, suppose that f and g have equal partial derivatives of order ≤ k at

p. Then (df) and (dg) have equal partial derivatives of order ≤ k − 1 at p, so by the

induction hypothesis we have (df)p ∼k−1 (dg)p. Therefore f ∼k g at p.

Corollary 3.2.2. f ∼k g at p iff the Taylor expansions of f and g up to (and

including) order k are identical at p

The composition of maps corresponds to the composition of jets in a nice way:

Lemma 3.2.3. Let U be an open subset of Rn and V an open subset of Rm. Let

f1, f2 : U → V and g1, g2 : V → Rl be smooth maps. If f1 ∼k f2 at p and g1 ∼k g2 at

f(p), then g1 · f1 ∼k g2 · f2 at p.

Proof. We proceed by induction on k. For k = 1, this is the chain rule:

d(g1 · f1)p = (dg1)f(p) · (df1)p = (dg2)f(p) · (df2)p = d(g2 · f2)p

Assume true for k−1. By hypothesis, (df1) ∼k−1 (df2) on TpU and (dg1) ∼k−1 (dg2)

on Tf(p)V . Applying the induction hypothesis point-wise, we have d(g1 · f1) ∼k−1

d(g2 · f2) on TpU . Therefore g1 · f1 ∼k g2 · f2 at p.

Proposition 3.2.1 and Lemma 3.2.3 are enough to prove that Jk(X,Y ) is a mani-

fold. The approach is a merely an elaboration on that taken with the tangent bundle.

First, we pull back charts on X × Y to charts on Jk(X,Y ) via the map α× β. Then

we verify that these latter charts are compatible, forming a smooth atlas.

31

We first need to establish some notation. Let Akn be vector space of polynomials

in n variables of degree ≤ k which have their constant term equal to zero. Choose

as coordinates for Akn the coefficients of the polynomials. Let Bk

n,m =⊕m

i=1 Akn and

define the map τ kn,m : C∞(Rn,Rm) → Bk

n,m to send each smooth map f = (f1, ..., fm)

to the coefficients (except the constant term) of the kth-order Taylor polynomial of

each fi in the obvious way.

We can now describe the aforementioned chart (W, η) of Jk(X, Y ) coming from

the charts (U,ϕ) of X and (V, ψ) of Y . We set W = (α × β)−1(U × V ) and define

the map η : W → ϕ(U)× ψ(V )×Bkn,m by

η(σ) =(ϕ(α(σ)), ψ(β(σ)), τ k(ψ · f · ϕ−1)

)

where f represents σ. Note that η is well-defined by Corollary 3.2.2.

To verify that such (W, η) form a smooth atlas on Jk(X, Y ), we must check that

these charts are compatible. Essentially, this reduces to the claim that a change of

basis on Rn smoothly sends Taylor series in the first basis to Taylor series in the

second basis. We now state our conclusions and refer the unconvinced reader to [5,

p.37].

Theorem 3.2.4. Let X and Y be smooth manifolds with n = dim X and m = dim Y .

Then

(1) Jk(X,Y ) is a smooth manifold with

dim Jk(X, Y ) = n + m + dim(Bkn,m)

(2) α : Jk(X,Y ) → X, β : Jk(X,Y ) → X, and α × β : Jk(X, Y ) → X × Y are

smooth submersions.

(3) If f : X → Y is smooth, then jkf : X → Jk(X, Y ) is smooth.

32

Since Jk(X,Y ) is a manifold, we can adapt the Transversality Lemma to a state-

ment about jets.

Proposition 3.2.5. Let X, B, and Y be smooth manifolds, W a submanifold of

Jk(X, Y ). Let G : X × B → Y be a smooth map and define Φ : X × B → Jk(X, Y )

by Φ(x, b) = jkGb(x). If Φ t W then {b ∈ B | jkGb t W} is a dense subset of B.

Proof. If G is smooth than Φ is smooth by part (3) of Theorem 3.2.4. Define j :

B → Jk(X, Y ) by j(b) = jkGb. Then Φ(x, b) = j(b)(x) and we are done by the

Transversality Lemma.

In §3.3, we will use the topology of Jk(X,Y ) to define a topology on C∞(X,Y ). In

§3.4, we will use Proposition 3.2.5 to prove the Thom Transversality Theorem. The

remainder of this section is devoted to further study of the differential structure on

Jk(X, Y ). In particular, we establish a collection of submanifolds of the jet bundle

that will be quite valuable in §4, §5, and §6.

Definition 3.2.2. Let σ ∈ J1(X,Y ) with source p. Let f represent σ.

(1) The rank of σ is defined to be the rank of (df)p.

(2) The corank of σ is defined to be the corank of (df)p.

(3) The singularity set of corank r is defined to be

Sr = {σ ∈ J1(X,Y ) | corank (σ) = r}

Note that if f and g both represent σ, then f ∼1 g and thus (df)p = (dg)p.

Therefore the rank and corank of 1-jets are well-defined. In order to prove that Sr is

a manifold, we must first do some linear algebra.

Let V and W be real vector spaces with n = dim V , m = dim W , and q =

min{n,m}. Define Lr(V, W ) = {S ∈ Hom(X,Y ) | corank S = r}.

33

Proposition 3.2.6. Lr(V,W ) is a submanifold of Hom(V,W ) with codim Lr(X, Y ) =

(n− q + r)(m− q + r).

Proof. Let S ∈ Lr(V, W ) and k = q − r = rank S. Choose bases on V and W

such that S takes the form

A B

C D

with A a k × k invertible matrix. Let U be

a coordinate nbhd of S in Hom(X,Y ) such that if S ′ =

A′ B′

C ′ D′

∈ U then A′

is invertible. We can choose such a U because the determinant map is continuous.

Define φ : U → Hom(Rn−k,Rm−k) by

A′ B′

C ′ D′

7→ D′ − C ′A′−1B. Fixing A′,

B′,and C ′ restricts φ to a translation. Therefore φ is a submersion. The following

lemma implies that Lr(V,W )⋂

U = φ−1(0). Then by Theorem 2.1.1, Lr(V, W ) is a

manifold and codim Lr(V, W ) = dim Hom(Rn−k,Rm−k) = (n− k)(m− k).

Lemma 3.2.7. Let S =

A B

C D

be an m × n matrix with A a k × k invertible

matrix. Then rank S = k iff D − CA−1B = 0.

Proof. The matrix

T =

Ik 0

−CA−1 Im−k

is an m×m invertible matrix. So

rank (S) = rank (TS) = rank

A B

0 D − CA−1B

This latter matrix has rank k iff D − CA−1B = 0

34

Theorem 3.2.8. Let X and Y be smooth manifolds with n = dim X, m = dim Y ,

and q = min{n,m}. Then Sr is a submanifold of J1(X,Y ) with codim Sr = (n− q +

r)(m− q + r).

Proof. Let σ ∈ Sr with source p and target q. Let U and V be coordinate nbhds of

p and q respectively. Then J1(X,Y )U×V∼= U × V × Hom(Rn,Rm) and under this

isomorphism Sr∼= U × V × Lr(Rn,Rm). Now apply Proposition 3.2.6.

3.3 The Whitney C∞ Topology

With the jet bundle in hand, we are ready to define a topology on C∞(X, Y ).

Definition 3.3.1. Let X and Y be smooth manifolds. Fix a non-negative integer k.

To each subset U of Jk(X,Y ), we associate a subset M(U) of C∞(X, Y ) defined by

M(U) = {f ∈ C∞(X, Y ) | jkf(X) ∈ U}

The Whitney Ck topology on C∞(X,Y ) is the topology whose basis is the family of

sets {M(U)} where U is an open subset of Jk(X,Y ). Denote by Wk the open subsets

in the Whitney Ck topology. The Whitney C∞ topology on C∞(X, Y ) is the topology

whose basis is W =⋃∞

k=0 Wk. We sometimes abbreviate this name to the C∞ topology.

Proposition 3.3.1. The Whitney Ck and C∞ topologies are well-defined.

Proof. The family of sets {M(U) | U ⊂ Jk(X,Y ) is open} forms a topological basis

because M(∅) = ∅, M(Jk(X,Y )) = C∞(X, Y ), and M(U)⋂

M(V ) = M(U⋂

V ). So

the Whitney Ck topology is well-defined.

To extend this result to the C∞ case, it suffices to verify that Wk ⊂ Wl when k ≤ l.

For each such k and l, define the map πlk : J l(X,Y ) → Jk(X,Y ) as follows: Let σ

35

be an l − jet with source p, and let f represent σ. Then πlk(σ) is the unique k-jet σ′

with source p such that f represents σ′. It is clear from Proposition 3.2.1 that πlk is

well-defined and continuous. Note M(U) = M((πlk)−1(U)) and if U is an open subset

of Jk(X, Y ) then (πlk)−1(U) is an open subset of J l(X, Y ). Therefore Wk ⊂ Wl.

In order to get a feel for this topology, we will exhibit a neighborhood basis of a

smooth map f . Let d be a metric on Jk(X, Y ) compatible with its topology (In §5we will see that such a metric exists). For each continuous map δ : X → R+, define

Bδ(f) = {g ∈ C∞(X, Y ) | d(jkf(x), jkg(x)) < δ(x) for every x ∈ X}

Proposition 3.3.2. The Bδ(f) form a neighborhood basis at f in the Ck topology on

C∞(X, Y ).

Proof. We divide the proof into three parts.

Part A: Bδ(f) is open.

Part B: If W is a nbhd of f ∈ C∞(X, Y ), then there exists δ : X → R+ such that

Bδ(f) ⊂ W .

Part C: Given continuous maps δ, γ : X → R+, there exists a continuous map

η : X → R+ such that Bδ(f)⋂

Bγ(f) = Bη(f).

Proof of Part A: Define the map ∆ : Jk(X, Y ) → R by σ 7→ δ(α(σ))−d(jkf(α(σ)), σ).

Then ∆ is continuous, U = ∆−1(0,∞) is open, and Bδ(f) = M(U).

Proof of Part B: Let W be a nbhd of f ∈ C∞(X, Y ). Then there exists an open

subset V of Jk(X, Y ) with f ∈ M(V ) ⊂ W . Define ν : X → R+ by

ν(x) = inf{d(σ, jkf(x)) | σ ∈ α−1(x)⋂

(Jk(X, Y )− V )}

with ν(x) = ∞ if α−1(x) ⊂ V . ν is well-defined because α−1(x)⋂

(Jk(X,Y ) − V )

is closed and jkf(x) ∈ V . Also ν is bounded below by a positive constant on any

36

compact subset of X. Thus we can construct a continuous map δ : X → R+ such

that δ(x) < ν(x) for every x ∈ X using a partition of unity (see §5.1 for the definition

of this tool). Then Bδ(f) ⊂ M(V ) ⊂ W .

Proof of Part C: Define η : X → R+ by η(x) = min{δ(x), γ(x)}. η is continuous

and it is clear from the definition of Bδ(f) that Bδ(f)⋂

Bγ(f) = Bη(f).

We may think of Bδ(f) as consisting of those smooth maps whose first k partial

derivatives are all δ-close to f . If X is compact, we can find a countable nbhd basis by

taking the collection of Bδn(f) where δn ≡ 1n. So in this case, the C∞ topology satisfies

the first axiom of countability, and one may prove straightforwardly that a sequence

of functions fn in C∞(X, Y ) converges to f (in the C∞ topology) iff jkfn converges

uniformly to jkf . One can further check that jk : C∞(X, Y ) → C∞(X, Jk(X,Y )) is

continuous with respect to the C∞ topology on the domain and codomain [5, p.46].

We will need two more topological notions in the sections that follow:

Definition 3.3.2. Let F be a topological space. A subspace G of F is residual if it

is the countable intersection of open dense subsets. F is a Baire space if all residual

subsets are dense.

Lemma 3.3.3 (Baire Lemma). Let X and Y be smooth manifolds. Then C∞(X, Y )

with the C∞ topology is a Baire space.

A proof of the Baire Lemma may be found in [5, p.44]. Note that, in general, the

properties open, dense, and residual are completely unordered in terms of strength.

However, the Baire Lemma implies that all residual subsets of C∞(X, Y ) are dense in

C∞ topology.

37

3.4 The Thom Transversality Theorem

We will need the following lemma:

Lemma 3.4.1. Let X and Y be smooth manifolds, W a submanifold of Y . Let

TW = {f ∈ C∞(X,Y ) | f t W}.

If W is closed then TW is an open subset of C∞(X, Y ) (in the C1, and thus, C∞,

topology).

Proof. Define a subset U of J1(X,Y ) as follows. Let σ ∈ J1(X,Y ) with source p ∈ X

and target q ∈ Y . Let f represent σ. Then σ ∈ U iff either

(i) q /∈ W

(ii) q ∈ W with TqY = TqW + (df)p(TpX).

It follows that j1f(p) ∈ U iff j1f tp W . Thus

TW = {f ∈ C∞(X,Y ) | j1f t W}

= {f ∈ C∞(X,Y ) | j1f(X) ⊂ U}

= M(U).

Recall that the collection {M(U) | U ⊂ J1(X, Y ) is open} forms a basis for the

C1 topology. Thus it suffices to show that U is open, or equivalently that V =

J1(X, Y )− U is closed.

Let σ1, σ2, ... be a convergent sequence with σi ∈ V for every i and limi→∞ σi = σ.

We will show that σ ∈ V . Let p = α(σ) and q = β(σ), and let f represent σ. Since

σi /∈ U and W is closed, we conclude that q = limi→∞β(σi) lies in W . Choose charts

(U,ϕ) at p in X and (V, ψ) at q in Y such that f(U) ⊂ V . Via these charts we may

38

assume that X = Rn, Y = Rm, and W = Rk ⊂ Rm. Let π : Rm → Rm/Rk = Rm−k

be projection. By Lemma 3.1.2, f t W iff π · f is a submersion at 0 iff π · (df)0 /∈ F

where

F = {A ∈ Hom(Rn,Rm) | rank A < m− k}.

Define the map

η : Rn ×W × Hom(Rn,Rm) ⊂ J1(Rn,Rm) → Hom(Rn,Rm−k)

by η(x,w,B) = π · B. η is continuous and F is closed (as follows from Proposition

3.2.6), so η−1(F ) is closed in Rn × W × Hom(Rn,Rm) which, in turn, is closed in

J1(Rn,Rm). Moreover V is precisely η−1(F ) since (x, y, (dg)0) ∈ V ⇔ g is not trans-

verse to W at 0 ⇔ η(x, y, g) = π · (dg)0 ∈ F . Since V is closed in the local situation,

we conclude that σ is in V .

Recall Proposition 3.1.6, in which we showed that we can make any map X → Y

transverse to a fixed submanifold of Y by an arbitrarily small translation of the map.

The Thom Transversality Theorem makes an even stronger claim: we can make the k-

jet of any map X → Y transverse to a fixed submanifold of Jk(X, Y ) by an arbitrarily

small perturbation of the map. A translation will not suffice here since it would leave

the partial derivatives of the map fixed. Instead, we will need to perturb the map

locally by a polynomial. To realize this local perturbation smoothly, we need the help

of cut-off functions. Namely, if U ⊂ X is open and V ⊂ X is compact with V ⊂ U ,

then there exists a smooth function ρ : X → R (a cut-off function) such that

ρ =

{1 on a neighborhood of V

0 off U

See [7, p.36] for a proof of this standard fact.

39

Theorem 3.4.2 (Thom Transversality Theorem). Let X and Y be smooth man-

ifolds and W a submanifold of Jk(X,Y ). Then the set

TW = {f ∈ C∞(X,Y ) | jkf t W}

is a residual subset of C∞(X, Y ) in the C∞ topology. Moreover, TW is open if W is

closed.

Proof. Our proof follows [5, p.54]. We must show that TW is the countable intersection

of open dense subsets. Choose a countable covering of W by open subsets W1,W2, ...

in W such that each Wr satisfies:

a) The closure of Wr in Jk(X, Y ) is contained in W .

b) W r is compact.

c) There exist coordinate neighborhoods Ur ⊂ X and Vr ⊂ Y such that π(Wr) ⊂Ur × Vr, where π : Jk(X, Y ) → X × Y is the projection α× β.

d) U r is compact.

We can do this by first collecting such an open nbhd at each point of W . Since W is

second countable, we can then extract a countable subcover. Now let

TWr = {f ∈ C∞(X, Y ) | jkf t W on W r}.

Then TW =⋂∞

r=1 TWr , so we have reduced the proof to showing that each TWr is open

and dense.

Open: Let Tr = {g ∈ C∞(X, Jk(X,Y )) | g t W on W r}. Tr is open by Lemma

3.4.1. jk : C∞(X, Y ) → C∞(X, Jk(X, Y )) is continuous. Therefore TWr = (jk)−1(Tr)

is open. Note that the same reasoning shows that TW is open if W is closed.

Dense: Choose charts ψ : U → Rn and η : V → Rm and smooth functions

ρ : Rn → [0, 1] ⊂ R and ρ′ : Rm → [0, 1] ⊂ R such that

40

ρ =

1 on a neighborhood of ψ · α(W r)

0 off ψ(Ur)ρ′ =

1 on a neighborhood of η · β(W r)

0 off η(Vr)

These cut-off functions exist because W r is compact.

Let B′ be the space of degree k polynomial maps Rn → Rm. For b ∈ B′, define

gb : X → Y by

gb(x) =

f(x) if x /∈ Ur or f(x) /∈ Vr.

η−1(ρ(ψ(x))ρ′(ηf(x))b(ψ(x)) + ηf(x)) otherwise.

By inspection the map (x, b) 7→ gb(x) is smooth.

Now define Φ : X × B′ → Jk(X, Y ) by Φ(x, b) = jkgb(x). We claim that we can

construct an open nbhd B of 0 in B′ such that Φ|X×B will be transverse to W on

some nbhd of W r. Assuming this, we can apply Proposition 3.2.5 on X×B to obtain

b1, b2, ... in B converging to 0 such that jkgbit W on W r. Since g0 = f and gb = f

off Ur, we have limi→∞ gbi= f . Therefore TWr is a dense subset of C∞(X, Y ).

We now construct a nbhd B of 0 in B′ such that Φ|B×X : B ×X → Jk(X,Y ) is

a local diffeomorphism and thus transverse to every submanifold of Jk(X,Y ). Let

ε = 12min{d(supp ρ,Rm − η(Vr)), d(ηβ(W r), (ρ

′)−1[0, 1))}. Since Φ(x, b) ∈ W r, we

have that x ∈ α(W r) and gb(x) ∈ β(W r). Then s = d(ηf(x), ηgb(x)) < ε because

ηgb(x) = ρ(ψ(x))ρ′(ηf(x))b(ψ(x)) + ηf(x).

So

s = |ρψ(x)ρ′ηf(x)bψ(x)|

≤ |bψ(x)| < ε if ψx ∈ supp ρ

= 0 if ψx /∈ supp ρ.

41

Now gb(x) is in β(W r), so using the definition of ε we conclude that ηf(x) is in the

interior of (ρ′)−1(1). Thus our choice of ε sufficiently limits the polynomial pertur-

bation b so that ηf(x) lies within the region on which ρ′ ≡ 1 when Φ(x, b) ∈ W r.

Recall as well that ρ ≡ 1 on a nbhd of ψα(W r). Taken together, we have that

ηgb(x) = bψ(x) + ηf(x) and that gb(x′) = η−1(bψ + ηf)(x′) for all x′ in a nbhd of

x. This argument also holds for all b′ in some nbhd of b. We can now define a

smooth inverse σ 7→ (x′, b′) of Φ near (x, b) as follows. For σ ∈ Jk(X,Y ) near Φ(x, b),

let x′ = α(σ), and let b′ be the unique polynomial map of degree ≤ k such that

σ = jk(η−1(b′ψ + ηf))(x′). Therefore, as desired, Φ|B×X is a local diffeomorphism at

(x, b).

We can immediately generalize the Thom Transversality Theorem to deal with

multiple submanifolds of the jet bundle.

Corollary 3.4.3. Let X and Y be smooth manifolds, {Wi} be a countable collection

of submanifolds of Jk(X, Y ). Let

TW = {f ∈ C∞(X, Y ) | jkf t Wi for every 1 ≤ i ≤ k}

Then TW is a residual subset of C∞(X,Y ) in the C∞ topology. If the number of Wi

is finite and each Wi is closed, then TW is open as well.

Before generalizing this result further in §3.5, we ought to at least mention a more

elementary transversality theorem which, though intuitive, suffers from not being

particularly useful.

Theorem 3.4.4 (Elementary Transversality Theorem). Let X and Y be smooth

manifolds, W a submanifold of Y . Then the set {f ∈ C∞(X, Y ) | f t W} is a residual

subset of C∞(X,Y ) in the C∞ topology. Moreover, this set is open if W is closed.

42

The Elementary Transversality Theorem can be proven directly by the same

method we used to prove the Thom Transversality Theorem. Alternatively, it can

be deduced rather quickly from Thom by noting that J0(X, Y ) = X × Y and

j0f(x) = (x, f(x)). We leave either approach to the industrious reader.

3.5 The Multijet Transversality Theorem

For the reader who fears the Thom Transversality Theorem to be insufficiently ab-

stract, we now generalize still one step further. In particular, we will need the Multijet

Transversality Theorem to prove claims about the generality of certain global prop-

erties of smooth maps, such as injectivity in §5.2.

Definition 3.5.1. Let X and Y be smooth manifolds. Define

Xs = X × · · · ×X (s times)

X(s) = {(x1, ..., xs) ∈ Xs | xi 6= xj for every 1 ≤ i < j ≤ s}

Define the multijet source map αs : (Jk(X, Y ))s → Xs to be α × · · · × α (s times).

Then

(1) Jks (X, Y ) = (αs)−1(Jk(X, Y )) is the s-fold k-jet bundle of maps from X to Y

or equivalently

(1′) Jks (X, Y ) is the multijet bundle of all s-tuples of k-jets with distinct sources.

Note that X(s) is an open subset of Xs and therefore a submanifold. In addition,

αs is continuous so Jks (X, Y ) is an open subset of (Jk(X, Y ))s and a submanifold as

well. Finally, we define jks : C∞(X,Y ) → C∞(X(s), Jk

s (X, Y )) by jks f(x1, ..., xs) =

(jkf(x1), ..., jkf(xs)).

43

Theorem 3.5.1 (Multijet Transversality Theorem). Let X and Y be smooth

manifolds, W a submanifold of Js. Then the set

TW = {f ∈ C∞(X,Y ) | jks f t W}

is a residual subset of C∞(X, Y ) in the C∞ topology.

Proof. The proof follows the same strategy as the Thom Transversality Theorem, by

proving a multijet analog to Lemma 3.4.1, then using the Transversality Lemma to

prove a multijet analog to Proposition 3.2.5. The only complication is that we must

now effect local polynomial perturbations of f on s regions simultaneously. This can

be done (with care) because the sources of an s-fold k-jet are distinct.

One can go further to show that if W is compact, than TW is open [5, p.57], but

we will not need this result.

In §1, §2, and §3, we developed powerful theorems on transversality.

In §4, §5, and §6, we use them to study the singularities of smooth maps.

Chapter 4

High to Low: Morse Theory

In §4.1, we give normal forms for non-degenerate critical points. In §4.2, we show that

non-degenerate critical points are the only generic singularities of smooth functions.

4.1 The Morse Lemma

Let X be a smooth manifold, f : X → R a smooth function. Since R is one dimen-

sional as a manifold, the derivative of f must have rank zero or one at each p ∈ X.

Thus a critical point p of f is simply a point for which all the partial derivatives of f

vanish. Relative to any coordinate system we have:

(∂f∂x1

)p

= ... =(

∂f∂xn

)p

= 0

However, not all critical points are created equal. The following tool encodes the

critical information that we will use to construct normal forms for the structure of

functions near most critical points.

Definition 4.1.1. Let f : X → R be a smooth function.

(1) The Hessian of f at p, with respect to local coordinates x1, ..., xn, is the matrix

Hp(f) =(

∂2f∂xi∂xj

)p

of second order partial derivatives:

44

45

(∂2f∂x2

1

)p

· · ·(

∂2f∂x1∂xn

)p

.... . .

...(

∂f2

∂xn∂x1

)p· · ·

(∂2f∂x2

n

)p

(2) A critical point p of f is degenerate if det Hp(f) = 0. Otherwise, p is non-

degenerate.

(3) The index of f at a non-degenerate critical point p is the maximum dimension

of a vector subspace of Rn on which Hp(f) is negative definite.

Remark 4.1.1. Hp(f) is negative definite on V if the corresponding bilinear form

H : Rn × Rn → R is negative definite, i.e. H(v, v) < 0 for every non-zero v ∈ V .

Equivalently, the index can be viewed as the number of negative eigenvalues of the

non-singular Hessian matrix.

Note that we have defined the Hessian of f at p in a way that depends on the

particular chart chosen at p. There also exists an invariant formulation of the Hessian

using the concept of intrinsic derivative [5, 64]. While we have avoided the latter

approach for simplicity, we must now do a little work to verify that the degeneracy

and index of a function at a point are well-defined notions.

Proposition 4.1.1. The degeneracy and index of f at p do not depend on the coor-

dinates chosen on X.

Proof. Let A = Hp(f) be the Hessian matrix with respect to the coordinates x1, ..., xn

given by a chart (U,ϕ) of X at p. Let ψ : ϕ(U) → ϕ(U) be a change of coordinates

defined by ψ(x1, ..., xn) = (y1(x1, ..., xn), ..., yn(x1, ..., xn)). Then the matrix P =

(dψ)0 is non-singular, and the matrix of the Hessian of f at p with respect to the

coordinates y1, ..., yn is given by (P−1)T AP−1. The latter claim is an exercise in

46

quadratic forms, namely that a change of coordinates replaces a quadratic form with

matrix A by a quadratic form with matrix BT AB, where B is non-singular. Clearly

A is singular iff BABT is singular. And if A is non-singular, than A and BABT have

the same index by Sylvester’s Law [1, p.245].

By the Submersion Lemma, a smooth function is locally equivalent at a regular

point to projection onto the first coordinate. The Morse Lemma provides normal

forms for the local behavior of smooth functions at non-degenerate critical points.

Theorem 4.1.2 (Morse Lemma). Let f : X → R be a smooth function, p ∈ X

a non-degenerate critical point of f , and λ the index of f at p. Then near p, f is

equivalent to the map (x1, ..., xn) 7→ f(p)− x21 − ...− x2

λ + x2λ+1 + ... + x2

n.

Our proof of the Morse Lemma fleshes out the sketch given by Milnor [10, p.6]

and will require the following calculus result.

Lemma 4.1.3. Let f be a smooth function on some convex region V ⊂ Rn, with

f(0) = 0. Then there exist smooth functions g1, ..., gn on V with

f(x1, ..., xn) = Σni=1xigi(x1, ..., xn)

and gi(0) = ∂f∂xi

(0) for every 1 ≤ i ≤ n.

Proof.

f(x1, ..., xn) =∫ 1

0dfdt

(x1t, ..., xnt)dt =∫ 1

0

(∑ni=0

∂f∂xi

(x1t, ..., xnt)xi

)dt

by the fundamental theorem of calculus and the chain rule. Note that convexity

guarantees that the above integral is defined. So it suffices to set

gi(x1, ..., xn) =∫ 1

0∂f∂xi

(x1t, ..., xnt)dt

47

where gi(0) = ∂f∂xi

(0) again follows from the fundamental theorem of calculus.

Proof of the Morse Lemma. In Part A, we will prove the existence of a change of

coordinates on the domain which yields the diagonalized quadratic form f(p)± x21 ±

...±x2n. In Part B we will show that the index of f(p)−x2

1− ...−x2λ + x2

λ+1 + ... + x2n

at 0 is λ.

Part A. We can assume without loss of generality that 0 = p = f(p) and X = Rn,

since we are only concerned with local equivalence. By Lemma 4.1.3, there exist

smooth functions g1, ..., gn on Rn with

f(x1, ..., xn) = Σni=1xigi(x1, ..., xn)

and gi(0) = ∂f∂xi

(0). Since 0 ∈ Rn is a critical point, we have ∂f∂xi

(0) = 0 for every

1 ≤ i ≤ n. Therefore we can apply Lemma 4.1.3 again, this time to each of the gi.

So there exist smooth functions hij, 1 ≤ i, j ≤ n, such that

gi(x1, ..., xn) = Σni=1xihij(x1, ..., xn)

Substitution gives

f(x1, ..., xn) = Σni,j=1xixjhij(x1, ..., xn)

Furthermore, we can assume hij = hji (otherwise replace each hij with 12(hij + hji)).

Differentiating gives ∂2f∂xi∂xj

(0) = 2hij(0), so the matrix (hij(0)) = (12( ∂2f

∂xi∂xj(0))). By

hypothesis, 0 is a non-degenerate critical point of f , so we conclude that (hij(0)) is

non-singular.

We now precede as in the proof of the diagonalization of quadratic forms [2, p.286].

Suppose by induction that there exist coordinates u1, ..., un on a neighborhood U1 of

0 such that

48

f(u1, ..., un) = ±u21 ± ...± u2

r−1 +∑n

i,j=r u1u2Hij(u1, ..., un)

on U1, where the Hij are smooth functions with Hij = Hji and the matrix (Hij(0))

non-singular. We have already established the base case r = 0.

For the induction step, we first show that we can make Hrr(0) 6= 0 by a non-

singular linear transformation on the last n− r + 1 coordinates. The proof works the

same for any r, so for simplicity let r = 1. If we have Hii(0) 6= 0 for some 1 ≤ i ≤ n

then we are done by transposing u1 and ui. Otherwise, since (Hij(0)) is non-singular,

there exists some Hij(0) 6= 0 with i 6= j. Through a pair of transpositions, we can

assume H11(0) = 0 and H12(0) = H21(0) 6= 0. We define a new set of coordinates

u′1, ..., u′n on U1 by

u′1 = 12(u1 + u2) u′2 = 1

2(u1 − u2) u′i = ui for i > 2

This linear transformation is invertible with inverse given by

u1 = (u′1 − u′2) u2 = (u′1 + u′2) ui = u′i for i > 2

Substituting in these new coordinates and regrouping terms, we have f(0) = Σni,j=1u

′iu′jH

′ij(0)

with H ′11(0) = H12(0) + H21(0) = 2H12 6= 0.

So without loss of generality we assume Hrr(0) > 0 (sending ur to −ur if neces-

sary). Then there exists a neighborhood U2 ⊂ U1 of 0 on which Hrr is positive. We

now define a new set of coordinates v1, ..., vn by

vi = ui for i 6= r.

vr =√

Hrr(u1, ..., un)[ur +

∑i>r uiHir(u1, ..., un)/Hrr(u1, ..., un)

]

Note vr is well-defined and smooth on U2. A simple calculation shows ∂vr

∂ur=√

Hrr

so ∂vr

∂ur(0) 6= 0. It follows from the Inverse Function Theorem that the change of

coordinates map φ defined by (u1, ..., un) 7→ (v1(u1, ..., un), ..., vn(u1, ..., un)) is a dif-

feomorphism in some sufficiently small neighborhood U3 ⊂ U2 of 0. Then

49

f = ±u21 ± ...± u2

r−1 +n∑

i,j=r

uiujHij

= ±u21 ± ...± u2

r−1 +

[u2

rHrr + 2ur

∑i>r

uiHri +∑i>r

u2i Hii/Hrr

]

+∑i>r

u2i (Hii −Hii/Hrr) +

∑

i,j>r,i6=j

uiujHij

The term in brackets is v2r so it is clear that we can choose smooth functions H ′

ij(v1, ..., vn)

for i > r so that

f(v1, ..., vn) =∑r

i=1±v2i +

∑ni,j>r vivjH

′ij(v1, ..., vn).

with H ′ij = H ′

ji. Furthermore, (H ′ij(0)) = ((dφ)−1

0 )T (Hij(0))(dφ)−10 is non-singular.

This completes the induction step and the first part of the proof.

Part B. Define g : Rn → R by g(x1, ..., xn) = g(p)− x21− ...− x2

λ + x2λ+1 + ... + x2

n.

Computing partial derivatives we have

Hp(g) =

−2

. . .

−2

2

. . .

2

The first λ basis vectors span a subspace V ⊂ R(n) on which Hp(g) is negative

definite, so the index of g at p is at least λ. The latter basis vectors span a subspace

W ⊂ Rn of dimension n − λ on which Hp(g) is positive definite. If there exists a

subspace V ′ of dimension greater than λ on which Hp(g) is positive definite, then V ′

50

and W would intersect nontrivially, a contradiction. Therefore, the index of g at p

equals λ.

A function f : X → R is called Morse if all of its critical points are non-degenerate.

Between the Submersion Lemma and the Morse Lemma, we have completely deter-

mined the local structure of Morse functions. In addition, we have the following

corollary.

Corollary 4.1.4. Non-degenerate critical points are isolated.

Proof. By the Morse Lemma, we need only consider the function f(x1, ..., xn) =

−x21 − ...− x2

λ + x2λ+1 + ... + x2

n. Clearly the partial derivatives ∂f∂xi

(p) = ±2xi do not

simultaneously vanish away from the origin.

4.2 Morse Functions are Generic

We have seen that Morse functions have simple local behavior, but this result would

be of little use if few functions satisfied the Morse property. In this section, we will

see that in fact almost all functions are Morse. First, we had better make “almost

all” precise.

Definition 4.2.1. A property P is generic if the set {f ∈ C∞(X,Y )|f has property P}is a residual subset of C∞(X, Y ).

The goal of this section is to show that the quality of being Morse is a generic

property of smooth functions. Our strategy will be to translate non-degeneracy into a

transversality condition on jets and apply the Thom Transversality Theorem. Recall

from §3.2 that Sr is the smooth submanifold of J1(X,R) consisting of those jets

51

which drop rank by r. For a smooth function f : X → R, only S0 and S1 can be

non-empty. Moreover, j1f : X → J1(X,R) maps critical points to S1 and regular

points to S0. The following proposition provides the key link between non-degeneracy

and transversality.

Lemma 4.2.1. Let f : X → R be a smooth function with a critical point at p. Then

p is non-degenerate iff j1f t S1 at p.

Proof. Degeneracy is a local property, so we assume X = Rn. Now J1(Rn,R) ∼=Rn × R × Hom(Rn,R). Let π : J1(Rn,R) → Hom(Rn,R) be the projection. By

Lemma 3.1.2, j1f t S1 at p iff π · j1f : Rn → Hom(Rn,R) is a submersion at p.

But π · j1f is simply the gradient of f . And ∇f : Rn → Rn is a submersion iff

(d∇f)p = Hp(f) is non-singular.

Theorem 4.2.2. Let X be a smooth manifold. The set of Morse functions is an open

dense subset of C∞(X,R) in the C∞ topology.

Proof. Note that codim S1 > 0 by Theorem 3.2.8, so Ss1 is a closed submanifold

of Jk(X,Y ). Thus the Thom Transversality Theorem and Baire Lemma imply that

{f ∈ C∞(X,R) | j1f t S1} is an open dense subset of C∞(X,R). By Lemma 4.2.1,

this set is exactly the set of Morse functions.

In fact, using the Multijet Transversality Theorem we can make a statement about

the critical values of smooth functions. We say that a function f has distinct critical

values if no two critical point of f have the same image.

Lemma 4.2.3. Let X be a smooth manifold. Then the set of smooth functions with

distinct critical values is a residual subset of C∞(X,R).

52

Proof. Let ∆Rn be the diagonal {(x, x) ∈ Rn × Rn}. We claim that the set S =

(S1 × S1)⋂

(β2)−1(∆R) is a submanifold of the multijet bundle J12 (X,R). Let U be

an open coordinate nbhd in X diffeomorphic to Rn. In local coordinates

J12 (U,R) ∼= (Rn × Rn −∆Rn)× (R× R)× Hom(Rn,R)2

S ∼= (Rn × Rn −∆Rn)× (∆R)× Hom(Rn, 0)2

so S is clearly a submanifold. Comparing the dimensions of the above factors, we see

codim S = 2n + 1.

Since S is a submanifold, we can apply the Multijet Transversality Theorem to

conclude that the set of maps f : X → Rn such that j12f t S is residual. Since

dim X(2) = 2n < 2n + 1 = codim S, by Proposition 3.1.1 we know that j12f t S

implies j12f(X(2))

⋂S = ∅. If p and q are distinct critical points of such an f , than

certainly j12f(p, q) = (j1f(p), j1f(q)) ∈ (S1×S1). Thus (j1f(p), j1f(q)) ∈ (β2)−1(∆R)

and we conclude f(p) 6= f(q).

The next proposition follows immediately from Lemma 4.2.3 and Theorem 4.2.2.

Proposition 4.2.4. Let X be a smooth manifold. The set of Morse functions with

distinct critical values is a residual subset of C∞(X,R).

In order to elucidate the importance of this theorem, we conclude the chapter with

a brief excursion into the theory of stable maps. Such maps are a central object of

study in higher dimensional Singularity Theory and Catastrophe Theory. In fact, the

latter field grew out of Thom’s and others complete classification of the singularities

of stable maps in dimension ≤ 4 (the so-called seven elementary catastrophes, two of

which are the topic of §6).

53

Definition 4.2.2. Let f : X → Y be a smooth map of manifolds. f is stable if the

equivalence class of f is open in C∞(X, Y ) with the C∞ topology.

Informally, f is stable if all nearby maps look like f . Note that if f is stable,

then all of its differentially invariant properties are unchanged by sufficiently small

perturbations of f .

Theorem 4.2.5. Let f : X → R be a smooth function. Then f is stable iff f is a

Morse function with distinct critical values.

We can easily prove the forward direction using Proposition 4.2.4, for it suffices

to note that any function equivalent to a Morse function with distinct critical points

is itself such a function. We invite the reader to fill in the details.

On the other hand, it is hard to show that Morse functions with distinct critical

values are stable. The difficulty is a more general one: there is no good direct method

for verifying the global stability of a map. V. Arnold and John Mather confronted

this problem by developing a local condition called infinitesimal stability which is

generally far easier to verify. At first glance, infinitesimal stability appears to be a

weaker condition than global stability, but in 1969 Mather succeeded in proving that

these properties are entirely equivalent [9]. This breakthrough precipitated a torrent

of major advances in the decade that followed.

Chapter 5

Low to High: The WhitneyEmbedding Theorem

In §5.1, we prove the that 1:1 proper immersions are embeddings and that proper

maps form a non-empty open subset of C∞(X,Rm). In §5.2, we show that there

are no generic singularities of maps from n-dimensional manifolds to m-dimensional

manifolds when m ≥ 2n. We then prove the Whitney Embedding Theorem.

5.1 Embeddings and Proper Maps

Definition 5.1.1. Let f : X → Y be a smooth map of manifolds. f is an embedding

of X into Y if f is a diffeomorphism of X onto a submanifold of Y . In this case, we

say f embeds X into Y .

The figures below illustrate the existence of 1:1 immersions which are not embed-

dings. In Figure 5.1, the real line is mapped into the plane to form a figure eight.

The image fails to be a submanifold at the central point. In Figure 5.1, the real line

is wrapped around the torus with an irrational ratio of equatorial to meridional loops

so as to be an injective. Yet the image fails to be a submanifold near all points.

54

55

Figure 5.1: [6, p.16]

Figure 5.2: [6, p.17]

It seems that we run into trouble when many points near infinity are mapped close

together. The following property gives us the additional control at infinity, preventing

us from packing in too many distant points into a compact set in the image.

Definition 5.1.2. Let f : F → G be a continuous maps of topological spaces. Then

f is proper if the preimage of every compact subset of G is compact in F .

Proposition 5.1.1. Let f : X → Y be a smooth map of manifolds. If f is a proper

1 : 1 immersion, then f embeds X into Y .

Proof. By the Immersion Lemma, if the image of a 1:1 immersion is a submanifold

then the inverse map is smooth. We show in two steps that if a 1:1 immersion is also

proper, then this image is in fact a submanifold.

Step 1: f is a homeomorphism onto f(X).

56

Step 2: f(X) is a submanifold.

Proof of Step 1: Let U be an open subset of X. We will show that f(U) is

open in f(X). Suppose by contradiction that f(U) is not open. Then there exists

a convergent sequence y1, y2, ... in f(X) − f(U) with limi→∞ yi = y ∈ f(U). Let

x1, x2, ... be the sequence of unique preimages of y1, y2, ... and let x be the preimage

of y. Then x ∈ U . On the other hand, {y, yi} is a compact subset of Y and f is

proper, so {xi} is compact in X. Thus we have a convergent subsequence xi1 , xi2 , ...

in X−U . Mapping this subsequence forward, we see that its limit must be x as well.

But since X − U is closed, we arrive at the contraction x ∈ X − U .

Proof of Step 2: Let p ∈ X, q = f(p). Since f is a homeomorphism onto its image,

we can transport each chart (U,ϕ) of X at p to a chart (f(U), ϕf−1) of f(X) at q.

By the Immersion Lemma, it is clear that the resulting differential structure on each

f(U) is locally compatible with that of Y . Thus f(X) is a submanifold of Y .

Consider the space C∞(X, Y ) of smooth maps from X to Y . If X is compact, then

all maps in C∞(X,Y ) are proper. However, if X is non-compact, then it is not imme-

diately obvious that a single proper smooth map exists. We will see that proper maps

always exist by first considering the space C∞(X,R). In order to construct a proper

map in C∞(X,R), we will need the following standard tool on smooth manifolds.

Definition 5.1.3. A partition of unity of a smooth manifold X is an open covering

{Ui} of X and a system of smooth functions ψi : X → R satisfying the following

conditions:

(1) For every x ∈ X we have 0 ≤ ψi ≤ 1.

(2) The support of ψi is contained in Ui.

(3) The covering is locally finite.

57

(4) For every x ∈ X we have∑

i ψi(x) = 1. Note that the sum is taken over all i

but is finite by (2) and (3).

A partition of unity {ψi, Ui} is subordinate to an open cover {Vj} of X if each Ui

is contained in some Vj.

The main theorem on partitions of unity is their existence for paracompact spaces:

Given any open cover {Vj} of a smooth manifold, there exists a partition of unity

{ψi, Ui} subordinate to {Vj} [7, p.32].

Lemma 5.1.2. Let X be a smooth manifold. Then there exists a smooth proper

function ρ : X → R.

Proof. Let {ψi, Ui}i=1,2,... be a partition of unity subordinate to a countable cover of

X by open sets with compact closure. Define ρ : X → R by ρ(x) =∑∞

i=1 iψi(x).

ρ is well-defined and smooth because partitions of unity are locally finite. Now if

ρ(x) ≤ j then by (4) we must have ψi(x) 6= 0 for some i ≤ j. Therefore ρ−1([−j, j]) ⊂⋃j

i=1{x|ψi(x) 6= 0}, the latter set having compact closure. So ρ−1([−j, j]) is a closed

subset of a compact set and therefore compact. To finish, note that every compact

subset of R is contained in [−j, j] for some j.

Having found a proper map in C∞(X,R), we now extend our result to C∞(X, Y ).

Proposition 5.1.3. Let X be a smooth manifold. Then the set of proper maps

X → Rm is a non-empty open subset of C∞(X, Y ).

Proof. Non-empty: By Lemma 5.1.2, there exists a smooth proper function ρ : X →R. Compose ρ with a linear injection R → Rm to obtain a smooth proper map

X → Rm.

58

Open: Let f : X → Rm be proper and let Vx = {y ∈ Rm | d(f(x), y) < 1}.Let V =

⋃x∈X x × VX in J0(X, Y ) = X × Rm. V is open because f is continuous.

Therefore M(V ) is an open in C∞(X,Y ) and clearly f ∈ M(V ). We will show that

if g ∈ M(V ) then g is proper. Let Bj = {y ∈ Rm | d(y, 0) ≤ j}. Then by the

triangle inequality we have g−1(Bj) ⊂ f−1(Bj+1). Thus g−1(Bj) is a closed subset of

a compact set, and thus compact. To finish, note that every compact subset of Rm is

contained in Bj for some j.

5.2 Proof of the Whitney Embedding Theorem

Let X be a smooth manifold of dimension n. The Whitney Embedding Theorem

states that X embeds into R2n+1. To see intuitively why 2n + 1 dimensions suffice,

trace a smooth curve in R3 with your fingertip. The drawn curve may intersect itself,

but with the slightest of adjustments we can push the curve off of itself entirely.

In this section, we will formalize this intuition using the tools developed in §3. But

first, it is interesting to note a more elementary approach which proves the Whitney

Embedding Theorem for manifolds that are already embedded in RN for some N . The

idea is to show that if N > 2n+1, then there exists a linear projection π : RN → RN−1

such that π · ρ : X → RN−1 is a 1:1 proper immersion. Thus by induction we can

reduce N to 2n + 1. This approach is taken by Guillemin and Pollack without loss

of generality because they define manifolds concretely as subsets of some RN at the

outset [6, p.51].

With our abstract definition, this approach reduces the Whitney Embedding The-

orem to showing that X can be realized as a subset of some RN . If X is compact,

59

we can use a cover of X by a finite number of charts to construct such an embedding

[12, p.223]. However, a constructive argument seems impossible in the case that X is

non-compact. Instead, we will prove existence using a more powerful approach based

on the Thom Transversality Theorem.

The following obvious lemma provides the key link between immersions X 7→ Y

and submanifolds of J1(X, Y ).

Lemma 5.2.1. f : X → Y is an immersion iff j1f(X)⋂ (⋃

r 6=0 Sr

)= ∅.

Proposition 5.2.2 (Whitney Immersion Theorem). Let X and Y be smooth

manifolds with dim Y ≥ 2 dim X. Then the set of immersions X → Y is an open

dense subset of C∞(X, Y ).

Proof. Open: By Theorem 3.2.8, S0 an open submanifold of J1(X,Y ). Thus, the set

of immersions X → Y , which equals M(S0), is open in C∞(X,Y ).

Dense: By Lemma 3.2.8, codim Sr = (n − q + r)(m − q + r) where n = dim X,

m = dim Y , and q = min{n,m}. Since dim Y ≥ 2 dim X, we have m ≥ 2n and q = n,

so for r ≥ 1 we have

codim Sr = r(m− n + r) ≥ m− n + r ≥ n + r > n.

Therefore, j1f t Sr for every r ≥ 1 iff j1f(X)⋂ (⋃

r 6=0 Sr

)= ∅ iff f is an immersion,

by Proposition 3.1.1 and Lemma 5.2.1. We are done by Corollary 3.4.3 to the Thom

Transversality Theorem.

Note where the dimension requirement is used: we must have m ≥ 2n in order

to enforce codim S1 = codim L1(Rn,Rm) > n and in turn apply Thom via Propo-

sition 3.1.1. We have shown that there are no generic singularities of maps from

60

n-dimensional manifolds to m-dimensional manifolds when m ≥ 2n. We now turn to

the global property of injectivity and formalize the intuition that most maps from a

small space to a large space are 1:1.

Lemma 5.2.3. Let X and Y be smooth manifolds with dim Y ≥ 2 dim X + 1. Then

the set of 1:1 maps X → Y is a residual subset of C∞(X, Y ).

Proof. Let n = dim X and m = dim Y . The diagonal ∆Y = {(y, y) ∈ Y × Y } is

diffeomorphic to Y and thus a submanifold of Y × Y dimension m. Define W =

(β2)−1(∆Y ), a submanifold of J02 (X, Y ) of codimension m by Theorem 2.1.2 and

Theorem 3.2.4. Let f : X → Y be a smooth map. Then f is 1:1 iff the image of

j02f : X(2) → J0

2 (X,Y ) does not intersect W . Since codim W = m > 2n = dim X(2),

by Proposition 3.1.1 we conclude that f is 1:1 iff j02f t W . We are done by the

Multijet Transversality Theorem.

Theorem 5.2.4. Let X and Y be smooth manifolds with dim Y ≥ 2 dim X +1. Then

the set of 1:1 immersions X → Y is a residual (and thus dense) subset of C∞(X,Y ).

Proof. The result is immediate from the Whitney Immersion Theorem, Lemma 5.2.3,

and the Baire Lemma.

In fact, the set of 1:1 immersions in these relative dimensions is open. The proof

of this fact does not require any new tools, but it is tedious [5, p.61]. At any rate, we

do not need it to prove our main theorem.

Theorem 5.2.5 (Whitney Embedding Theorem). Let X be a smooth manifold

of dimension n. Then X embeds into R2n+1.

61

Proof. The set of proper maps X → R2n+1 is non-empty and open by Proposition

5.1.3. The set of 1:1 immersions X → R2n+1 is dense by Theorem 5.2.4. Therefore

there exists a proper 1:1 immersion X → R2n+1, which is an embedding by Proposition

5.1.1.

For the curious, in these relative dimensions the stable maps are exactly the 1:1

immersions [5, p.81].

Whitney proved Theorems 5.2.2 and 5.2.5 in 1936 [17]. Eight years later, using

more difficult techniques in self-intersection theory and algebraic topology, Whitney

was able to improve the dimensions of each theorem by one: Any mapping of an

n-manifold (n ≥ 2) into R2n−1 may be perturbed slightly to an immersion [19], and

any n-manifold may be embedded in R2n [18]. At the time, Whitney remarked,“It

is a highly difficult problem to see if the imbedding and immersion theorems of the

preceding paper and the present one can be improved upon” [19, p.1]. However, as

algebraic topology evolved, it became clear that the first obstructions to immersing

an n-manifold X are the Stiefel-Whitney characteristic classes ωi(X) of the stable

normal bundle. In 1960, W. Massey proved that ωi(X) = 0 for i > n − ν(n), where

ν(n) is the number of ones in the base-2 expansion of n (so n = 2i1 + · · ·+ 2iν(n) with

i1 < · · · < ν(n)). This result is best possible since ωn−ν(n)(RP2i1 × · · ·RP2iν(n)

) 6= 0.

In particular, RP2i1 × · · · × RP2iν(n)

cannot be immersed in RN for N < n− ν(n).

It was thus conjectured that any n-manifold can be immersed in Rn−ν(n). By 1963

E. H. Brown and F.P. Beterson had strengthened Massey’s algebraic results, and in

1977 they proposed a program for proving this conjecture. In 1985, Cohen completed

this program [3].

Chapter 6

Two to Two: Maps from the Planeto the Plane

In §6.1 we prove that smooth maps (of 2-manifolds) with only folds and cusps form

an open dense set. We then give a normal form for folds. In §6.2 we give a normal

form for simple cusps.

6.1 Folds

In §4, we established normal forms for a generic set of maps from an n-manifold

down to R. Then in §5, we showed that almost all maps of an n-manifold into a 2n-

manifold have no singularities at all. We now turn our attention to the next simplest

case: maps from the plane to the plane. We will see that, generically, the singularities

of such maps consist of general fold curves with isolated cusp points.

Recall the singularity set Sr ⊂ Jk(X,Y ) of k-jets of corank r. In Theorem 3.2.8,

we saw that Sr is a submanifold of codimension (n−q+r)(m−q+r), where n = dim X,

m = dim Y , and q = min{n,m}. For a smooth map f , let Sr(f) = (jkf)−1(Sr) ⊂ X

denote set of the points of X at which the k-jet of f has corank r. If jkf t Sr then

62

63

Sr(f) is a submanifold X of codimension (n− q + r)(m− q + r) as well. Here we are

using Theorems 3.2.4 and 3.1.3.

Now let f : X → Y be a map of 2-manifolds with jkf t S1. Since we are

concerned with local structure of f at singularities, we may take X = Y = R2. From

the above remarks, we compute

dim J1(R2,R2) = dim(R2 × R2 ×B12,2) = 10

codim S1 = 1 =⇒ dimS1 = 9

codim S1(f) = 1 =⇒ dimS1(f) = 1.

Thus S1(f) is a collection of smooth curves in X, called the the general fold of f . At

each p ∈ S1(f), exactly one of the following two situations must occur.{

(a) TpS1(f)⊕

Ker(df)p = TpX;

(b) TpS1(f) = Ker(df)p.

Definition 6.1.1. Let f : X → Y be a map of 2-manifolds with j1f t S1. Then

p ∈ S1(f) satisfying (a) is called a fold point of f . And p ∈ S1(f) satisfying (b) is

called a cusp point of f .

The property that f has only folds and cusps is generic.

Proposition 6.1.1. Let X and Y be smooth 2-manifolds. Then the set of maps

X → Y with only fold and cusp singularities is an open dense subset of C∞(X, Y ).

Proof. We show that if f satisfies j1f t S1 and j1f t S2, then all critical points of

f are folds and cusps. By Thom and the Baire Lemma, such f form an open dense

subset of C∞(X, Y ) (noting that S1 and S2 are closed submanifold).

Suppose j1f t S2 and j1f t S1. A critical point p of f is either in S1(f) or S2(f).

But a non-empty S2(f) would be a submanifold of codimension 4 in X. So p is in

S1(f) and, thus, a fold or cusp.

64

We now give a normal form for the structure of a fold singularity.

Theorem 6.1.2. At a fold point, f is locally equivalent to the map (x1, x2) 7→(x1,±x2

2).

Proof. Let p be a fold point of f . Then j1f t S1 and we can assume p = f(p) = 0.

Consider the restriction map f |S1(f) from the general fold of f to Y . As a fold point,

0 satisfies T0S1(f)⊕

Ker(df)0 = T0X. f has rank 1 at p, so (df)0 must map T0S1(f)

injectively into T0Y . Thus d(f |S1(f))0 = (df)0|T0S1(f) is injective, and we conclude

f |S1(f) is an immersion at 0. By the Immersion Lemma, the local image of S1(f) is

a smooth curve (this will not be the case at a cusp). In fact, since f |S1(f) is a local

diffeomorphism onto its image, we have that f is locally equivalent to a smooth map

(x1, x2) 7→ (x1, g(x1, x2)). Computing (df)0, it is clear that S1(f) is exactly the set on

which ∂g∂x2

vanishes. Thus g(0) = ∂g∂x2

(0) = 0. Applying Lemma 4.1.3 twice as in the

proof of the Morse Lemma gives g(x1, x2) = x22h(x1, x2) for some smooth function h.

If h(0) = 0, then a simple computation shows that f would have the same 2-jet

at 0 as the map f given by (x1, x2) 7→ (x1, 0). Now the condition that j1f t0 S1 is a

condition on j1f(0) and (dj1f)0, i.e. on the 2-jet of f at 0. So we would have j1f t0 S1

as well and thus a submanifold S1(f) of X of codimension 1. This contradicts the

fact that f has rank 1 at every point of X.

Thus we have g(x1, x2) = ±x22h(x1, x2) with h(0) > 0, and the following is valid

change of coordinates on some nbhd of 0.

x1 = x1 x2 = x22

√h(x1, x2)

In these coordinates, f is the map (x1, x2) 7→ (x1,±x22).

Figure 6.1 illustrates why we call such a point a fold. We can visualize the map

65

(x1, x2) 7→ (x1, x22) by first mapping R2 to a parabolic cylinder in R3 via (x1, x2) 7→

(x1, x2, x22) and then projecting onto the (x1, x3) plane. The projection creases the

parabolic cylinder along the general fold S1(f).

Figure 6.1: [5, p.89]

6.2 Cusps

Let f : X → Y be a map of 2-manifolds with j1f t S1. Recall that a cusp point p of f

satisfies the cusp condition TpS1(f) = Ker(df)p. We would like to find a normal form

for cusps. This case is much more difficult than that of folds in §6.1. In fact, without

a further restriction on p there are infinitely many locally non-equivalent forms that f

could take at p. Fortunately, all but one of these forms almost never occur. In order

to distinguish the generic case, consider a non-vanishing vector field ξ along S1(f)

(near p) such that at each point of S1(f) ξ is in the kernel of (df). This is always

possible locally. By the cusp condition, ξ is tangent to S1(f) at p. The nature of the

cusp at p depends on what order of contact ξ has with S1(f) at p. To be precise, let

66

k be a smooth function on X with k = 0 on S1(f) and (dk)p 6= 0. Since k = 0 on

S1(f) and ξ is tangent to S1(f) at p, the function (dk)(ξ) : S1(f) → R has a zero at

p.

Definition 6.2.1. We say p is a simple cusp if this zero is a simple zero.

Theorem 6.2.1. At a simple cusp, f is locally equivalent to the map (x1, x2) 7→x1, x

31 + x1x2.

As with fold points, we can visualize simple cusps via a map to R3 and a projection:

Figure 6.2: [5, p.147]

67

The general fold is a parabola, and the image of the cusp point is indeed a cusp

of the image of the general fold. Whitney’s original proof of Theorem 6.2.1 involves

many lemmas followed by six pages of coordinate changes [20]. This approach, though

successful for the cusp, degenerates into an intractable mess for higher dimensional

singularities. In order to prove Theorem 6.2.1 in a clean and generalizable fashion, we

will call upon a powerful extension of the Weierstrass Preparation Theorem of complex

analysis (though it will be expressed in the language of commutative algebra). It is

perhaps not entirely surprising that complex analysis comes into play, as a major

obstacle to establishing normal forms for smooth real maps is that these maps are

not necessarily real analytic (we cannot replace them with their Taylor series). This

problem evaporates over the complex numbers, as smooth complex functions are

necessarily complex analytic.

We will need some terminology. Let X be a smooth manifold. By C∞p (X) we

denote the ring of smooth germs of functions on X at p. C∞p (X) is a local ring with

maximal ideal Mp(X) consisting of germs taking the value 0 at p. Note that R ∼=C∞p (X)/Mp(X). Thus if A is a finite-dimensional C∞p (X)-module, then A/Mp(X)A is

a real vector space.

Theorem 6.2.2 (Generalized Malgrange Preparation Theorem). Let X and

Y be smooth manifolds and f : X → Y a smooth map with f(p) = q. Let A be a

finitely generated C∞p (X)-module. Then A is a finitely generated C∞q (Y )-module iff

A/Mq(Y )A is a finite dimensional vector space over R.

Our proof of Theorem 6.2.1 will use Theorem 6.2.2 in the form of a more easily

applicable corollary, though we will not deduce the corollary here. Define inductively

a sequence of ideals Mkp (X) in C∞p (X) by letting M1

p (X) be Mp(X), and Mkp (X) be the

68

vector space generated by germs of the form fg with f in Mp(X) and g in Mk−1p (X).

Using Lemma 4.1.3, it is not hard to show by induction that Mk0 (X) consists precisely

of those germs of smooth functions f whose Taylor series at 0 begin with terms of

degree k.

Corollary 6.2.3. If the projections of e1, ..., ek form a spanning set of vectors in the

vector space A/(Mk+1p (X)A+Mq(Y )A), then e1, ..., ek form a set of generators for A

as a C∞q (Y )-module.

Proof of Theorem 6.2.1. Our proof follows that of Golubitsky and Guillemin [5, p.147].

Since f has rank 1 at p, through a linear change of coordinates on the domain we

can assume ∂f1

∂x16= 0 and then apply the Immersion Theorem to force f into the

form (x1, h(x1, x2)) with h smooth. We can also assume (df)0 =

1 0

0 0

in this

coordinate system, i.e. ∂h∂x1

= ∂h∂x2

6= 0.

We now show that d( ∂h∂x2

)0 6= 0. For suppose otherwise, i.e. that

∂∂x1

∂h∂x2

(0) = ∂∂x2

∂h∂x2

(0) = 0.

Then f has the same 2-jet as the map (x1, x2) 7→ (x1, δx2) where δ = 12

∂2h∂x2

1(0). As in

the proof of Theorem 6.1.2, we note that the latter map is of rank 1 everywhere and

thus its 1-jet is not transverse to S1 at 0. But this condition depends only on the

2-jet of the map, contradicting that j1f t0 S1.

As in the fold case, S1(f) is defined by the equation ∂h∂x2

= 0, and so at each point

of S1(X) the kernel of (df) is spanned by ∂h∂x2

. Thus ∂h∂x2

meets the requirement of

our vector field ξ in the definition of simple cusp, while ∂h∂x2

suffices for our function

k because ∂h∂x2

(0) = 0 and d( ∂h∂x2

)0 6= 0. The cusp condition is ∂2h∂x2

2(0) = 0, while the

simple cusp condition is ∂3h∂x3

2(0) 6= 0. Therefore, at the origin, we have

69

h = ∂h∂x2

= ∂2h∂x2

2= 0 and ∂3h

∂x326= 0.

We are now in a position to apply the corollary to the Malgrange Preparation Theo-

rem.

Recall that f is given by f(x1, x2) = (x1, h(x1, x2)). Via f , C∞0 (R2) becomes a

module over itself: a · b(x1, x2) = a(f(x1, x2))b(x1, x2) where a is in the ring C∞0 (R2)

and b is in the module C∞0 (R2). By Corollary 6.2.3, this module is generated by 1, x2,

and x22 if the vector space C∞0 (R2)/(M4

0 (R2) + (x1, h)) is generated by 1, x2, and x22.

(In the corollary’s notation, A = C∞0 (R2), Mp(X) = (x1, h), and Mq(X) = (x1, x2)).

The conditions on h guarantee that (x1, h) ⊃ M30 (R2). (This step makes evident why

it is necessary for the cusp to be simple in order to apply our corollary and arrive at

a simple normal form). Thus M30 (R2) ⊃ (M4

0 (R2) + (x1, h)) and the the vector space

C∞0 (R2)/(M40 (R2) + (x1, h)) is indeed spanned by 1, x2, and x2

2.

Since our module is generated by 1, x2, and x22, we can express x3

2 as

x32 = 3a2(x1, h)x2

2 + a1(x1, h)x2 + a0(x1, h).

where the ai are smooth functions of (y1, y2) = (x1, h) vanishing at 0. We can rewrite

this equation in the form

(x2 − a)3 + b(x2 − a) = c

with a = a2, and b and c new functions of (y1, y2) vanishing at 0. If we set x1 = 0

in this equation, the left hand side takes the form x32 + ..., the dots indicating terms

of order > 3 in x2. Since h(0, x2) = x32 + ... as well, the right and left hand sides

of the equation can only be equal if ∂c∂y2

(y1, y2) 6= 0 at 0. Now note that the leading

term of the Taylor series of h is a non-zero multiple of x1x2; so comparing the linear

terms on either side of the equation gives ∂b∂y1

6= 0 at 0, while comparing the quadratic

70

terms gives ∂c∂y1

= 0 at 0. Taken together, these results show that the following are

legitimate coordinate changes:

x1 = b(x1, h)

x2 = x2 − a(x1, h)

y1 = b(y1, y2)

y2 = c(y1, y2)

In these coordinates f takes the normal form (x1, x2) 7→ (x1, x31 + x1x2).

It is evident from Figure 6.2 that simple cusps are isolated. More rigorously, we

saw that with h = x32 − x1x2, the cusp condition on f = (x1, h) requires ∂2h

∂x22(0) = 0,

while the simple cusp condition requires ∂3h∂x3

2(0) 6= 0. Thus f has no other cusps in a

sufficiently small nbhd of a simple cusp.

With more effort, we could extend Proposition 6.1.1 to the following claim: the

set of smooth maps with only folds and simple cusps is open and dense. We will

refrain however, as a proper attack would lead us into the territory of higher order

analogs of the singularity sets Sr. In any case, the reader who has made it this far

ought to be convinced of the truth of the claim.

Bibliography

[1] Michael Artin, Algebra, Prentice-Hall, Upper Saddle River, New Jersey, 1991.

[2] Garrett Birkhoff and Saunders Mac Lane, A Survey of Modern Algebra, 4th Ed.,

MacMillan Publishing Co., New York, 1979. (First Ed. 1941).

[3] Ralph L. Cohen, The Immersion Conjecture for Differentiable Manifolds, Annals

of Math., 2nd Ser., Vol. 122, No. 2 (Sep., 1985), 237–328.

[4] C. G. Gibson, Singular Points of Smooth Mappings, Pitman Publishing, London,

1979.

[5] Martin Golubitsky and Victor Guillemin, Stable Mappings and Their Singulari-

ties, Springer-Verlag, New York, 1973. (Second Printing 1980).

[6] Victor Guillemin and Alan Pollack, Differential Topology, Prentice Hall, Engle-

wood Cliffs, New Jersey, 1974.

[7] Serge Lang, Differential Manifolds, Addison Wesley Publishing Co., Reading,

Massachusetts, 1972.

[8] Yung-Chen Lu, Singularity Theory and an Introduction to Catastrophe Theory,

Springer-Verlag, New York, 1976.

[9] John N. Mather, Stability of C∞ Mappings: II. Infinitesimal Stability Implies

Stability, Annals of Math., 2nd Ser., Vol. 89, No. 2 (Mar., 1969), 254–291.

71

72

[10] John Milnor, Morse Theory, Princeton University Press, Princeton, New Jersey,

1963. (Second Printing 1965).

[11] , Topology from the Differential Viewpoint, University of Virgina Press,

Charlottesville, Virginia, 1965. (Fourth Printing 1976).

[12] James R. Munkres, Topology: A First Course, Prentice-Hall, Englewood Cliffs,

New Jersey, 1975.

[13] D. J. Saunders, The Geometry of Jet Bundles, London Math. Soc. Lecture Note

Series 142, Cambridge University Press, Cambridge, 1989.

[14] Michael Spivak, Calculus on Manifolds, Addison-Wesley Publishing Co., New

York, 1965.

[15] Shlomo Sternberg, Lectures on Differential Geometry, Prentice-Hall, Englewood

Cliffs, New Jersey, 1964.

[16] Rene Thom, Structural Stability and Morphogenesis, W. A. Benjamin, Inc.,

Reading, Massachusetts, 1975.

[17] Hassler Whitney, Differentiable Manifolds, Annals of Math., 2nd Ser., Vol. 37,

No. 3 (Jul., 1963), 645–680.

[18] , The Self-Intersection of a Smooth n-Manifold in 2n-Space, Annals of

Math., 2nd Ser., Vol. 45, No. 2 (Apr., 1944), 220–246.

[19] , The Singularities of a Smooth n-Manifold in 2n − 1-Space, Annals of

Math., 2nd Ser., Vol. 45, No. 2 (Apr., 1944), 247–293.

[20] , On Singularities of Mappings of Euclidean Spaces. I. Mappings of the

Plane into the Plane, Annals of Math., 2nd Ser., Vol. 62, No. 3 (Nov., 1955),

374–410.